Datasets:
happy8825
/
anet_ret_train_global_rzen_sft

Dataset card Data Studio Files
xet
Community
1
messages
listlengths
2
2
videos
listlengths
5
5
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man is shown standing in front of a blue background as many harmonicas are shown and he starts playing them. He is then shown speaking into the camera and going into detail with regards to the products features. He shows off how many he has in his coat and continues talking. He is then shown playing another harmonica as he gives instruction on it. He speaks some more and shows his harmonica again as the video goes to an end title screen that reads Howcast original\".\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a man playing a saxophone, with shots alternating between him playing and speaking about lessons. It promotes saxophone lessons but doesn't match the harmonica content.\n\n[2] A man in front of a blue background showcases harmonicas, playing them and discussing their features. The video includes detailed demonstrations and ends with a \"Howcast original\" screen, closely matching the query.\n\n[3] Features a man discussing and playing a saxophone. It doesn't involve harmonicas or match the query content.\n\n[4] The video revolves around paintball gear, with scenes of outdoor play and discussion about masks. It is unrelated to the harmonica theme.\n\n[5] Shows a man in front of a blue background playing and demonstrating the harmonica, consistent with showing and explaining features. It ends with \"Howcast original,\" aligning with the query.\n</content>\n\n<contrast>\n- [2] is more closely aligned with the query as it involves detailed explanations and demonstrations of harmonicas, whereas [5] also matches but includes additional guitar elements.\n- [5] includes harmonica details and instruction within a similar setup, making it relevant, whereas [3] and [1] focus on saxophone content which doesn’t match the harmonica query.\n- [4] is unrelated due to its focus on paintball equipment instead of musical instruments.\n</contrast>\n\n<summary>\n1st=[2] (Direct match with harmonica features and instruction)\n2nd=[5] (Close match, focuses on harmonica with guitar)\n3rd=[3] (Focuses on saxophone, not harmonica)\n4th=[1] (Saxophone-focused, no harmonica)\n5th=[4] (Unrelated paintball content)\n</summary>\n</think>\n<answer> [2] > [5] > [3] > [1] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_CB0yYGVNcOk.mp4", "activitynet/videos/v_RULNhPrXnfA.mp4", "activitynet/videos/v_nDbLtdY66dA.mp4", "activitynet/videos/v_oO6jZR9Aijc.mp4", "activitynet/videos/v_PKdQIdYuXo8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man uses a large plastic blue bat. He swings it at a pinata shaped like spongebob squarepants. He misses several times, and keeps swinging.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a young boy swinging a bat at a pinata, but it's not a Spongebob pinata and he is missing more than hitting. The bat is not blue.\n\n[2] Features a Spongebob pinata with children present, but no large blue bat is evident. Various children are interacting with the pinata.\n\n[3] A man with a white shirt and red bandana swings a bat at a blue-green pinata. The bat is visible but not blue, and the pinata is not Spongebob.\n\n[4] This has an indoor setting with a pinata being hit. The pinata is colorful but not Spongebob, and the bat is not blue.\n\n[5] A man with a white shirt uses a large blue bat to swing at a Spongebob squarepants pinata. He seems to miss in a few frames and keeps swinging.\n</content>\n\n<contrast>\n- [5] matches the query as it features a man with a large blue bat swinging at a Spongebob pinata, missing a few times, whereas others don't meet all these criteria.\n- [2] has the Spongebob pinata but lacks the man with a blue bat.\n- [3] shows a man swinging a bat but not at a Spongebob pinata.\n- [1] involves a pinata and missing swings but neither the bat nor the pinata are correct.\n- [4] lacks the specific pinata and blue bat detail.\n</contrast>\n\n<summary>\n1st=[5] (matches query with blue bat and Spongebob pinata)\n2nd=[2] (Spongebob pinata is present, lacks blue bat)\n3rd=[3] (man with a bat hitting a pinata, but not Spongebob)\n4th=[1] (missing attempts with pinata but neither bat nor pinata are correct)\n5th=[4] (least related without specific objects from query)\n</summary>\n</think>\n<answer> [5] > [2] > [3] > [1] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_ac1DjqY4xHs.mp4", "activitynet/videos/v_3fg_CB3rBSk.mp4", "activitynet/videos/v_e_X0K2t8API.mp4", "activitynet/videos/v_CCL8kqQMCRw.mp4", "activitynet/videos/v_LFeoVokXAFY.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Someone is brushing their teeth with an electric toothbrush. They wear braces so they make sure to scrub all around very well. You can see the tongue moving around never really being comfortable inside of the mouth. The man continues to brush, toothpaste spewing down his lips a bit .\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The images show a man shaving his beard with an electric razor. The focus is on the shaving process, and there is no visible toothbrush or tooth brushing action.\n\n[2] A man is brushing his teeth with a manual toothbrush. The images show his mouth and teeth, but there are no braces visible. Toothpaste is present, but the tongue movement and discomfort are not evident.\n\n[3] The sequence features two individuals holding toothbrushes and toothpaste. The focus is more on the interaction between them and playful actions rather than detailed brushing techniques or the presence of braces.\n\n[4] A close-up view of someone brushing their teeth with a toothbrush. The individual is wearing braces, and there is significant toothpaste visible around the mouth. The focus is on thorough scrubbing around braces, aligning well with the query.\n\n[5] The video appears to depict a humorous or staged scene about brushing teeth. A man uses a toothbrush, but there is no focus on braces or detailed brushing technique as described in the query.\n</content>\n\n<contrast>\n- [4] is ranked higher than [2] because it shows braces and detailed brushing with visible toothpaste, matching the query closely, whereas [2] lacks braces.\n- [2] ranks above [5] due to more relevant brushing action, whereas [5] seems staged without specific details on braces.\n- [5] is preferred over [3] as it at least involves brushing, whereas [3] involves playful actions not focused on the brushing process.\n- [3] is above [1] because toothbrushes and toothpaste are present, whereas [1] features unrelated shaving action.\n</contrast>\n\n<summary>\n1st=[4] (includes braces, detailed brushing)\n2nd=[2] (brushing focus, lacks braces)\n3rd=[5] (involves brushing, lacks detail)\n4th=[3] (toothbrush present, not focused)\n5th=[1] (unrelated shaving action)\n</summary>\n</think>\n<answer> [4] > [2] > [5] > [3] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_Ktxr4ZiqO3o.mp4", "activitynet/videos/v_8l7SuE4_lCk.mp4", "activitynet/videos/v_Ia3Aa4AO-_E.mp4", "activitynet/videos/v_5kIl4TspdR0.mp4", "activitynet/videos/v_ZW3SKczQzJU.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man is seen playing a set of bagpipes while people behind him push around bicycles. He is seen playing again while more people walk in and out of frame with children and dogs. Some people stop to hear him play and he stops to look to the camera.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows a person playing an accordion on a city street. People are seen walking by, and some stop briefly.\n\n[2] Depicts a man playing bagpipes in a kilt with a crowd behind him. People are walking, some with bicycles.\n\n[3] Shows a bagpiper in traditional attire playing next to a child. It occurs in an outdoor area with onlookers.\n\n[4] Displays a person playing bagpipes in a park. Few people are present, no bicycles visible.\n\n[5] Shows a large group of bagpipers in a coordinated performance. The setting is indoors, with no bicycles.\n\n</content>\n\n<contrast>\n- [2] is most relevant because a man plays bagpipes in front of people with bicycles, fitting the query closely.\n- [3] involves bagpipes and people but lacks bicycles and crowd dynamics described in the query.\n- [4] features bagpipes but is in a park setting with minimal crowd interaction.\n- [5] involves bagpipes but is an indoor performance and lacks individual focus or the outdoor dynamic.\n- [1] involves a street performer but uses an accordion and not bagpipes, differing greatly from the query.\n\n</contrast>\n\n<summary>\n1st=[2] (involves bagpipes, bicycles, and people, closely matching the query)\n2nd=[3] (involves bagpipes and people, but lacks bicycles and crowd as described)\n3rd=[4] (features bagpipes but lacks relevant crowd dynamics and bicycles)\n4th=[5] (features bagpipes but in an irrelevant indoor setting)\n5th=[1] (features accordion rather than bagpipes, making it least relevant)\n</summary>\n</think>\n<answer> [2] > [3] > [4] > [5] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_UusDcf16TVQ.mp4", "activitynet/videos/v_-G8N5FxjZbw.mp4", "activitynet/videos/v_qmOYKPkURiw.mp4", "activitynet/videos/v_0dgmG3h9RLA.mp4", "activitynet/videos/v_sYAGVE3luRo.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"An unseen person navigates several Amazon dot com pages on exercise equipment. An Amazon page showing an exercise machine alone is shown, followed by another image of a Amazon page showing a woman using an exercise machine along with a price and information about the equipment. Another page showing a eliptical trainer is shown in which an active computer mouse is navigating the review for the item at which point another model elliptical trainer on another page is shown.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video focuses on a person using and adjusting an exercise machine. There's no evidence of Amazon pages or online navigation.\n\n[2] This video showcases multiple screenshots of Amazon product pages for an elliptical trainer. It begins with a standalone image of the equipment, followed by a model using it, price details, and review interactions, matching the query sequence closely.\n\n[3] The video features close-ups and demonstrations of an exercise machine in use, with branding and specifications highlighted. No Amazon page browsing is depicted.\n\n[4] The video includes various angles of an exercise machine, with people using it, and detailed shots of features, but does not involve any online page navigation.\n\n[5] Similar to [4], it shows the exercise machine with users and specific features, but without any Amazon page interactions.\n\n</content>\n\n<contrast>\n- [2] has clear relevance due to explicit displays of Amazon pages featuring exercise equipment, whereas [1], [3], [4], and [5] focus on physical use and features of machines without online context.\n- [3] and [4]/[5] focus on product demonstration and features, whereas [2] includes navigation elements related to the query.\n- [1] depicts physical engagement with equipment, mismatching the query about online navigation.\n\n</contrast>\n\n<summary>\n1st=[2] (direct match with Amazon page browsing sequence)\n2nd=[3] (detailed product demonstration but lacks Amazon navigation)\n3rd=[4] (focused on product features, not Amazon navigation)\n4th=[5] (similar to [4], no online element)\n5th=[1] (only shows exercise use, no Amazon context)\n</summary>\n</think>\n<answer> [2] > [3] > [4] > [5] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_WElK913B33U.mp4", "activitynet/videos/v_43j1BAyP6Q0.mp4", "activitynet/videos/v_sV6JEbmqDUw.mp4", "activitynet/videos/v_f6NpnKJZJE0.mp4", "activitynet/videos/v_Yr2unlRQZa0.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman in a pink shirt is riding a cycle bike. She is then doing push ups on the ground. She does sit ups on the ground.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The sequence shows a woman in a pink outfit doing sit-ups on the ground, with no evidence of cycling or push-ups.\n\n[2] This sequence features a woman in a pink shirt cycling, then doing push-ups, and finally sit-ups on the ground. Fits the query perfectly in terms of actions.\n\n[3] Contains interviews and cycling but lacks the specific sequence of activities matching the query.\n\n[4] Shows a woman engaging in cycling and ground exercises, but her shirt is not pink and the sequence is incomplete.\n\n[5] Displays cycling and sit-ups, but her shirt is not pink, and the sequence is not coherent with the query.\n</content>\n\n<contrast>\n- [2] matches the query perfectly with sequential actions of cycling, push-ups, and sit-ups, whereas [1] only shows sit-ups.\n- [2] contains the correct attire and sequence, whereas [3] lacks ground exercises fitting the query.\n- [2] has complete coverage with the right shirt color, whereas [4] and [5] miss either the correct shirt color or sequence completeness.\n</contrast>\n\n<summary>\n1st=[2] fits the query exactly with the correct sequence and attire.\n2nd=[4] contains some relevant actions but lacks the pink shirt.\n3rd=[5] shows relevant actions but misaligns on attire and sequence.\n4th=[1] only shows sit-ups, missing a comprehensive match.\n5th=[3] lacks necessary physical activities and sequence.\n</summary>\n</think>\n<answer> [2] > [4] > [5] > [1] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_x8yuq2i72eo.mp4", "activitynet/videos/v_MGF7I9aCPPM.mp4", "activitynet/videos/v_QxsELAtsSWE.mp4", "activitynet/videos/v_UsexiHSzIQI.mp4", "activitynet/videos/v_BhxKjirVF4Y.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Two men are standing on the side of the court holding basketball. The men are giving interview, talking to the camera, while one man is talking the other man is listening, then vice versa. The man went to the basket dribble to the ball, the men ran to the basket and shot the ball.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows two coaches, but the sequence lacks clear interview scenes. There is basketball practice, but not the described specific interactions or shots.\n\n[2] Two men are seen talking and holding a basketball at the side of the court. They take turns speaking to the camera. One of them dribbles, runs, and shoots the ball, aligning well with the query.\n\n[3] Similar setup to [2] with two men taking turns speaking on the court. They dribble and practice shooting, matching the query requirements.\n\n[4] This video features an individual talking on the court without a clear interview interaction or specific dribbling and shooting transition.\n\n[5] Displays a single person interacting with the camera, dribbling, and shooting basketballs. However, it lacks the dual interview dynamic.\n</content>\n\n<contrast>\n- [2] beats [3] because it shows better sequential coverage of speaking and playing.\n- [3] beats [1] because it contains relevant court interactions and actions described in the query, while [1] does not.\n- [1] beats [4] because [1] at least involves basketball practice.\n- [4] beats [5] because it involves more comprehensive coverage in relation to the query’s demands, though lacking interaction.\n</contrast>\n\n<summary>\n1st=[2] (Aligns perfectly with query: talking, dribbling, and shooting.)\n2nd=[3] (Shows interview and action but less coherently than [2].)\n3rd=[1] (Includes basketball practice but lacks clear dual interaction.)\n4th=[4] (Has court presence but lacks the query’s interactive elements.)\n5th=[5] (Focuses on individual actions without the described interaction.)\n</summary>\n</think>\n<answer> [2] > [3] > [1] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_9029wVyGkvg.mp4", "activitynet/videos/v_eW5W8dVB_I4.mp4", "activitynet/videos/v_UwZbkuVsPno.mp4", "activitynet/videos/v_p8h3G0Tblqc.mp4", "activitynet/videos/v_iYxj8a1TPYk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man is seen running down the street holding a box and speaking to another. The man then plays a game with a small girl and bows after each game. The two continue to play with the man laughing at the camera and him giving her a kiss in the end.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The sequence shows a man interacting with a small girl, playing a game together where the man bows and appears to have a friendly, playful interaction. It aligns well with the query.\n[2] The sequence focuses on an older man and a boy playing a tabletop game with some interaction, but it doesn't fully match the query's description of running and interaction with a girl.\n[3] This sequence shows a street event with people dancing and socializing, unrelated to the query.\n[4] The sequence is about a boy hitting a piñata at a party, not related to the query activity or interaction.\n[5] The sequence involves a child setting up and playing with a hopscotch mat outdoors, not connected to the query involving a man's interaction with a girl.\n</content>\n\n<contrast>\n- [1] is more specific to the query due to the man's playful interaction with a girl, whereas [2] involves an older man and a boy with no running or bowing.\n- [2] has some interaction with a game, whereas [3] involves no interaction fitting the query.\n- [4] involves a stationary activity (hitting piñata), whereas [5] involves kids setting up and playing a game, both not involving the described interaction.\n</contrast>\n\n<summary>\n1st=[1] (clearly meets the query: man interacting with a girl, playing, bowing)\n2nd=[2] (involves a game, but wrong participants and missing elements)\n3rd=[4] (a game/activity is present, wrong kind and missing girl)\n4th=[5] (outdoor play setup with child, missing man and described elements)\n5th=[3] (social event, no game or specific interaction fitting the query)\n</summary>\n</think>\n<answer> [1] > [2] > [4] > [5] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_nKnsvy7-wAI.mp4", "activitynet/videos/v_n6A8YtBAJX0.mp4", "activitynet/videos/v_i1llwxnI_Vo.mp4", "activitynet/videos/v_AmW-O5MmJHk.mp4", "activitynet/videos/v_X4IE65LtDzQ.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"The starting list for the men's pole vault event is displayed over the background of the competition area. Multiple vaulters attempt to vault, one after another, some successfully and some not. The camera cuts to a man wiping his face with a towel, then a quick shot of the last vault attempt, then a quick shot of another man walking away from the camera. Another man in white and blue vaults and is successful. The camera cuts to four other competitors watching. The camera cuts to a man in white and blue celebrating while holding a flag against his black. The man who celebrated is being interviewed. A different competitor, wearing red, is being interviewed.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This sequence features a men's pole vault competition, highlighting multiple vaulters, successful attempts, and celebrations. It also includes interviews, aligning well with the query details.\n\n[2] The frames depict a weightlifting competition, focusing on athletes in action and audience reactions, not related to pole vaulting.\n\n[3] This sequence covers a track event, focusing on a single athlete, but not pole vaulting.\n\n[4] The frames show a men's pole vault competition, including attempts, successes, audience reactions, and celebrations, closely matching the query.\n\n[5] This sequence captures various pole vault attempts in outdoor settings, showing different athletes and heights, generally related to the query but lacks detailed narrative coherence.\n\n</content>\n\n<contrast>\n- [1] beats [4] because it includes interviews and detailed narrative flow matching the query.\n- [4] beats [5] because of clearer focus on specific athletes and events with celebrations.\n- [5] beats [3] because it involves pole vaulting, whereas [3] involves a track event unrelated to pole vaulting.\n- [3] beats [2] because it involves track events, which are more related to the query than weightlifting.\n</contrast>\n\n<summary>\n1st=[1] (matches query with details of attempt, celebration, interview)\n2nd=[4] (good coverage of pole vaulting events with celebrations)\n3rd=[5] (covers pole vault but lacks narrative details)\n4th=[3] (track event not pole vaulting)\n5th=[2] (unrelated weightlifting event)\n</summary>\n</think>\n<answer> [1] > [4] > [5] > [3] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_pjF6uBZHVRE.mp4", "activitynet/videos/v_GXEuuWhEMCY.mp4", "activitynet/videos/v_nmUJ2GfVkKY.mp4", "activitynet/videos/v_RQO61KXeBos.mp4", "activitynet/videos/v_OH5UXwa3F7E.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman is sitting at a table using a hookah while another woman makes silly faces and gestures behind her. They speak animatedly with the person behind the camera. There is a close up of the items on the table in front of them. The woman with the camera turns it to face herself before aiming it back toward the other women. One of the women coughs as a cloud of smoke hits her in the face.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A young man uses a hookah, with close-ups of the device and smoke blowing. No women or silly gestures present. \n[2] Two young men use a hookah in close-up, blow smoke, but no women or table interaction is visible. The scene lacks silliness or women interacting with a camera.\n[3] Two women are involved with a hookah at a table. One woman makes silly gestures. There are close-ups of the hookah and a brief camera turn. A cloud of smoke causes coughing.\n[4] A group uses a hookah, with a man and women involved, but there is no table setting, silly gestures, or camera turning; it doesn't fully match the detailed query.\n[5] A hookah is used at a table with two women talking, but no silly gestures or coughing scenes. The camera does not appear to turn to the person capturing the scene.\n</content>\n\n<contrast>\n- [3] includes two women, a hookah, silly gestures, table interaction, and a camera turn, whereas [5] lacks silly gestures and camera turning.\n- [5] features women and a hookah at a table, whereas [4] lacks the specified silliness and camera interaction.\n- [4] involves multiple people with a hookah, whereas [2] focuses on two men without the necessary context.\n- [2] lacks the interaction elements and women seen in [1].\n</contrast>\n\n<summary>\n1st=[3] (Matches query with women, hookah, gestures, table, and camera action)\n2nd=[5] (Women and hookah at a table but lacking gestures and camera turning)\n3rd=[4] (Group with a hookah; lacks necessary elements like silliness or camera action)\n4th=[2] (Two men with a hookah; lacks women and humorous context)\n5th=[1] (Man with a hookah; no women, gestures, or interaction)\n</summary>\n</think>\n<answer> [3] > [5] > [4] > [2] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_1D7iddqIIhM.mp4", "activitynet/videos/v_lgWH94ea3-U.mp4", "activitynet/videos/v_oEd7WT4mVSE.mp4", "activitynet/videos/v_SbX8KDWMhio.mp4", "activitynet/videos/v_Vbkq1ldn3A8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"woman is walking wearing a bikini and eating a chocolate is snowy day. people are in sidewalk watching the girl walks. woman is in a restaurant wearing a bikini and two aiter are watching her. woman walks in a airport eating a restaurant.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a woman applying sunscreen in different settings, mostly in a bikini, suggesting a beach setting but not involving snow or eating chocolate. It mostly focuses on sunscreen application tips.\n[2] This video involves athletes preparing and participating in a sports track event, unrelated to the query about walking in a bikini, snow, or restaurants.\n[3] Includes surfing, biking, and other outdoor activities, with some scenes involving swimwear but not in a snowy or restaurant setting.\n[4] Displays activities in a snowy environment, featuring skiing or snowboarding, but doesn't include a woman in a bikini or eating chocolate.\n[5] Features a woman in a bikini walking on a snowy day, eating chocolate in various settings including a restaurant and an airport, which aligns with the query.\n</content>\n\n<contrast>\n- [5] is the best match because it directly depicts the woman in a bikini walking in snowy settings, eating chocolate, and covers different scenarios including a restaurant and an airport.\n- [1] contains scenes with women in bikinis but lacks snow and chocolate references.\n- [3] involves swimwear but focuses on surfing and outdoor activities rather than the specific actions in the query.\n- [4] shows snow but no bikini or chocolate elements, focusing instead on winter sports.\n- [2] is entirely about sports events and not related to the query scenario.\n\n</contrast>\n\n<summary>\n1st=[5] (matches all query details – bikini, chocolate, snow, restaurant, airport)\n2nd=[1] (includes bikini-related content but lacks snow, chocolate, or specific settings)\n3rd=[3] (features swimwear but in unrelated activities like surfing)\n4th=[4] (has snow but focuses on winter sports, not fitting the query style)\n5th=[2] (sports track events, unrelated to query elements)\n</summary>\n</think>\n<answer> [5] > [1] > [3] > [4] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_ITGlho0dw-k.mp4", "activitynet/videos/v_fruimvo90vA.mp4", "activitynet/videos/v_YPXeQVi39nU.mp4", "activitynet/videos/v_EHVkkRI7mBQ.mp4", "activitynet/videos/v_jUPbimyUndA.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A large group of people are at a carnival playing in their own bumper cars and a few of them bump into one another and come to a stop. They laugh and begin to pull their bumper cars away from one another. The people start driving around again where they all continue to either bump into one another or just drive around. Eventually all the bumper cars come to a complete stop and the ride is over.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Many people are driving bumper cars indoors. There is some interaction as cars bump and then separate. The sequence captures moments of driving and stopping, consistent with the query of a bumper car ride.\n\n[2] This shows bumper cars in an outdoor carnival setting. People in cars bump into each other and continue driving, which aligns well with the query. The interaction between cars and stops is evident.\n\n[3] Features a night scene with bumper cars under lights. People drive and bump into each other. The energetic atmosphere and ending with cars stopping match the query details.\n\n[4] Shows bumper cars with people driving and interacting. There is less visible interaction compared to others, focusing more on movement. The setting is dim, but some stopping is visible.\n\n[5] Displays a dimly lit scene with fewer bumper cars. Interaction between cars and concluding stops are visible, but the overall activity is less intense.\n</content>\n\n<contrast>\n- [2] is better than [3] because it clearly shows bumper cars in a carnival and has more visible interactions.\n- [3] is better than [1] because the nighttime setting with clear bumps and final stops captures the carnival feel vividly.\n- [1] is better than [4] because it shows more interactions and fits the dynamic aspect of the query better.\n- [4] is better than [5] because although dimly lit, it shows more car interactions and people.\n- [5] is less dynamic than others, having fewer bumper car interactions.\n</contrast>\n\n<summary>\n1st=[2] (Clear setting and interactions well-suited to the query)\n2nd=[3] (Dynamic, nighttime atmosphere with visible stops)\n3rd=[1] (Good interactions and sequence of driving/stopping)\n4th=[4] (Less interaction, dimly lit, but some stopping is evident)\n5th=[5] (Least dynamic, fewer interactions, less intense activity)\n</summary>\n</think>\n<answer> [2] > [3] > [1] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_tQ9yPS_BJHQ.mp4", "activitynet/videos/v_bX8QXK9mjKA.mp4", "activitynet/videos/v_SiqB6_9kNQU.mp4", "activitynet/videos/v_Fp8e9dbWeSg.mp4", "activitynet/videos/v_jx7_0MRlScw.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A boy in an orange shirt is standing in a room. He starts punching a punching bag. The punching bag moves in the air.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A person in black practice martial arts with nunchucks in a gym. There is no punching bag involved.\n[2] An adult in gray works with a red punching bag, performing movements but not outright punching.\n[3] A boy in an orange shirt performs a jump rope routine outdoors. No punching bag is involved.\n[4] A person in an orange shirt throws a discus in an athletic field. No punching bag is involved.\n[5] A boy in an orange shirt punches a black punching bag, making it move. This aligns well with the query.\n</content>\n\n<contrast>\n- [5] aligns perfectly because it shows a boy in an orange shirt punching a bag, making it move, whereas [3] lacks any punching bag interaction.\n- [3] lacks a punching bag and involves jump rope, whereas [1] focuses on nunchuck practice not involving a punching bag.\n- [1] involves nunchuck practices, with no orange shirt or punching bag as per the query, whereas [2] features bag interaction but involves an adult.\n- [2] shows bag interaction but does not involve a boy; [4] involves a discus throw without any bag, matching the query even less.\n</contrast>\n\n<summary>\n1st=[5] (a boy in an orange shirt punches a punching bag making it move)\n2nd=[2] (bag interaction, but with an adult, not a boy)\n3rd=[1] (gym training but with nunchucks, no bag or orange shirt)\n4th=[3] (orange shirt, no bag, jump rope activity)\n5th=[4] (no bag or orange shirt, involves discus throw)\n</summary>\n</think>\n<answer> [5] > [2] > [1] > [3] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_tilmCzSf1oc.mp4", "activitynet/videos/v_fQDj7pnGdVY.mp4", "activitynet/videos/v_-NndIs9BaS4.mp4", "activitynet/videos/v_4E40jmmc2N8.mp4", "activitynet/videos/v_gU81ZXdYh7o.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman is sitting down wearing green sunglasses. She is talking on the cell phone. She is riding in a car. She is laying on a chair getting her foot tattooed. She is showing her foot tattoo. She is sitting outside talking again.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A woman is walking outside, stops to talk, and later is seen getting a tattoo. There is no indication of her wearing green sunglasses or riding in a car.\n[2] A woman with green sunglasses is shown talking outside, riding in a car while talking on the phone, getting a foot tattoo, and then showing her tattoo.\n[3] Focuses on men in a tattoo parlor with no women present or action related to the query.\n[4] A woman is in various indoor settings, but there's no tattoo insight or any outdoor or car scenes.\n[5] A person with sunglasses is indoors talking, and there are scenes of tattooing, but not all actions outlined in the query are visible.\n</content>\n\n<contrast>\n- [2] is most relevant as it sequentially shows the woman performing actions directly matching the query, including wearing green sunglasses, riding in a car, getting and showing a foot tattoo.\n- [1] involves tattoo scenes but lacks key elements like green sunglasses and riding in a car.\n- [5] features tattooing and sunglasses but not related to a woman or outdoor/car scenes.\n- [3] and [4] have no direct correlation with the query, focusing on scenarios unrelated to a woman’s tattoo and sunglasses.\n</contrast>\n\n<summary>\n1st=[2] (matches all components: green sunglasses, riding in a car, tattoo)\n2nd=[1] (shows tattoo but lacks sunglasses and car)\n3rd=[5] (sunglasses and tattoo but not a woman or outdoor context)\n4th=[4] (indoor setting, lacks relevant actions)\n5th=[3] (male subjects, tattoo context but different focus)\n</summary>\n</think>\n<answer> [2] > [1] > [5] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_AyYpsQ4xOo8.mp4", "activitynet/videos/v_kWPvAxLzGw8.mp4", "activitynet/videos/v_-8FS4LbXAgE.mp4", "activitynet/videos/v_vgO3mzJGbSI.mp4", "activitynet/videos/v_PtoVL69LEh4.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A boy on a team throws a basketball at other team members. Others join in, some falling as they are hit with the ball.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows a dodgeball game in a gym, with players actively throwing and dodging balls. The action is lively and includes people falling as they are hit, which aligns well with the query.\n\n[2] The video features children playing in a trampoline park with multiple balls involved. There is dynamic movement and jumping, with some players getting hit and falling. This matches the query context of a team throwing balls.\n\n[3] Shows a gym setting where children are playing with balls, but the emphasis seems more on playing than actively hitting others, with minimal falling action.\n\n[4] Displays people engaging in a ball game in a gymnasium, but the action is less intense with few instances of players falling or being hit.\n\n[5] Depicts a basketball practice session with players dribbling and passing a basketball. This does not match the dodgeball-like context of the query.\n</content>\n\n<contrast>\n- [1] beats [2] because it shows a clear dodgeball game with players specifically throwing and being hit by balls.\n- [2] beats [3] because it includes more evident instances of players getting hit and falling.\n- [3] beats [4] because there is more ball-related action in a gym setting, even if less focused on hitting.\n- [4] beats [5] because it involves a ball game setting, although not aligned with the query of throwing at team members.\n- [5] is last as it involves basketball, not fitting the dodgeball scenario.\n</contrast>\n\n<summary>\n1st=[1] (clear dodgeball game, players hit and fall)\n2nd=[2] (trampoline dodgeball-like play)\n3rd=[3] (gym setting with ball play, less hitting)\n4th=[4] (ball game, less intense action)\n5th=[5] (basketball, unrelated to query)\n</summary>\n</think>\n<answer> [1] > [2] > [3] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_PLgcXxX3A2A.mp4", "activitynet/videos/v_wvlyV067oOU.mp4", "activitynet/videos/v_pHZogKGkc9o.mp4", "activitynet/videos/v_UySzNwvkQKQ.mp4", "activitynet/videos/v_B7Q1egUKmZg.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Two young men are shown running back and fourth down a court hitting a tennis ball. They keep hitting the ball against the wall and chasing the ball around. One pauses and hits the ball on his racket and again throws it for the boys to play. One falls down on the ground towards the end and the other helps him up.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Two young men are playing squash, running back and forth, hitting the ball against the wall. One falls, and the other helps him up, aligning well with the query about a similar tennis wall-play scenario.\n[2] A young boy and an adult playing in a squash court do not match the query's context of two young men.\n[3] Instructional video on squash for children, focused on technique, not relevant to the query’s scenario.\n[4] Two players hitting the ball against the wall, but lacks the specific details of the query, such as one helping the other up.\n[5] Two men playing in a squash court, lacking the specific interactions described in the query like falling and helping each other.\n</content>\n\n<contrast>\n- [1] best matches the query as it captures two young men playing and helping each other after one falls, whereas [4] lacks the helping aspect.\n- [5] involves squash play but misses the fall and assist, while [2] features a child and adult, not two young men.\n- [3] focuses on teaching rather than play, with children instead of young men.\n</contrast>\n\n<summary>\n1st=[1] (best matches with continuous play and helping interaction)\n2nd=[4] (features players hitting ball but no fall or help)\n3rd=[5] (squash play with men but lacks fall and assist)\n4th=[2] (young child and adult, mismatch in age and context)\n5th=[3] (instructional, with children, not relevant scenario)\n</summary>\n</think>\n<answer> [1] > [4] > [5] > [2] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_rggwT0-G5sk.mp4", "activitynet/videos/v_P3BkKdkf43I.mp4", "activitynet/videos/v_q6bl7g5JGng.mp4", "activitynet/videos/v_jIYRVlrXJgc.mp4", "activitynet/videos/v_aWz9H0jTv3c.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A young girl is seen sitting in a chair with a person standing next to her. The person next to her then piercing one ear followed by the other. The person rubs lotion on the piercings afterwards.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A young girl is sitting at a table, but there's no evidence of an ear piercing taking place. The focus is on the girl's nails being attended to.\n[2] A young girl is visibly having her ears pierced while sitting, followed by cleaning of the area. This matches the ear-piercing process described in the query.\n[3] A young girl is shown having both ears pierced while seated. The sequence includes piercing each ear and applying lotion, directly aligning with the query.\n[4] The video features a young girl having her ears pierced, but the lotion application isn't clearly visible.\n[5] A young girl is getting her ears pierced, but the presence of lotion application isn't shown clearly.\n</content>\n\n<contrast>\n- [3] clearly shows the complete sequence of ear piercing followed by lotion application, whereas [2] shows piercing and some post-care, but less clearly.\n- [2] shows the piercing and has some evidence of post-care, whereas [4] lacks clear lotion application visibility.\n- [4] shows the ear piercing activity, whereas [5] also shows piercing but not lotion application.\n- [1] doesn’t show any ear piercing activity, focusing instead on nail care, whereas the others show ear piercing.\n</contrast>\n\n<summary>\n1st=[3] (completes sequence including lotion application; highest query match)\n2nd=[2] (clear ear piercing but less focus on lotion)\n3rd=[4] (piercing visible, lotion less clear)\n4th=[5] (piercing visible but lotion application not shown)\n5th=[1] (not related to ear piercing)\n</summary>\n</think>\n<answer> [3] > [2] > [4] > [5] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_KCxa27MM7Cg.mp4", "activitynet/videos/v_yyvc2OSPx1Y.mp4", "activitynet/videos/v_69IsHpmRyfk.mp4", "activitynet/videos/v_Jbl0E75cx8o.mp4", "activitynet/videos/v_PQgydNnSggw.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man in a suit is sitting behind a desk. People are playing lacrosse on a field of grass. A person in a yellow uniform is standing in front of a net blocking balls.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video begins with a man in a suit standing behind a desk with a sports backdrop, followed by scenes of people playing a field sport involving sticks, possibly lacrosse or hurling, on a grass field.\n[2] This video features a bagpiper on a sports field, followed by several frames of a hurling match with players in helmets playing on grass. There are no scenes with a man in a suit.\n[3] Starts with a man in a suit at a desk in a newsroom setting, cutting to a field hockey team on grass. A person in a yellow uniform is blocking balls in front of a net.\n[4] Displays a lacrosse game with players in yellow and green uniforms, but no man in a suit or someone in a yellow uniform blocking balls specifically seen.\n[5] Shows a field with various activities involving people in white and some players in a sporting scenario, but no man in a suit or clear ball blocking is visible.\n</content>\n\n<contrast>\n- [3] matches all parts of the query: man in a suit at a desk, lacrosse, and yellow uniformed person blocking balls, whereas [1] lacks the yellow uniformed ball blocking.\n- [1] features a man in a suit and a similar sport but doesn't show a person in a yellow uniform blocking balls, whereas [2] has none of these aspects.\n- [4] shows lacrosse but lacks both the man in a suit and protection actions, whereas [3] includes these elements.\n- [5] has partial sports activity but lacks a man in a suit and ball-blocking action, whereas [1] does include a man in a suit and similar sports activities.\n</contrast>\n\n<summary>\n1st=[3] (includes all query elements: man in a suit, lacrosse, and yellow uniformed player blocking balls)\n2nd=[1] (man in a suit and field sport but no yellow uniformed player blocking)\n3rd=[4] (lacrosse coverage but lacks other query elements)\n4th=[2] (hurling and field sport but missing the man in a suit and blocking scene)\n5th=[5] (partial sports activity, lacks suit and blocking scene)\n</summary>\n</think>\n<answer> [3] > [1] > [4] > [2] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_xww6n1FoJIM.mp4", "activitynet/videos/v_agZNSscDJww.mp4", "activitynet/videos/v_t0ajvfx6dgA.mp4", "activitynet/videos/v_sJUUm0n4cXA.mp4", "activitynet/videos/v_XazKuBawFCM.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man was sitting inside a room. He is holding a bowl of noodles and broth. He is drinking the broth from a bowl.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The man is outside, sitting under a tent, holding a bottle, and drinking. There's no bowl or noodles visible, which doesn’t match the query.\n[2] Shows a man smoking a hookah in a cafe. He is not holding a bowl or drinking broth, thus irrelevant to the query.\n[3] A person is in a bathroom, touching water and washing his face. No noodles or drinking action is visible.\n[4] A man is holding a bowl, sitting inside a room, and drinking from it. It matches the query closely as he appears to be consuming broth.\n[5] Two men are sitting in a cafe with paper cups, likely containing coffee. No bowl or noodles are involved.\n</content>\n\n<contrast>\n- [4] is relevant because the man holds a bowl and drinks from it, matching the action described in the query, whereas [1], [2], [3], and [5] do not match these specific actions.\n- [1] shows drinking but lacks a bowl with noodles, whereas [4] shows the correct action with a bowl.\n- [5] involves drinking from a cup, which doesn't align with the query of drinking broth from a bowl.\n</contrast>\n\n<summary>\n1st=[4] (Shows a man with a bowl, drinking, matches the query)\n2nd=[1] (Shows drinking action but lacks bowl with noodles)\n3rd=[5] (Drinking present but from a cup, not bowl)\n4th=[2] (Completely unrelated activity, not drinking broth)\n5th=[3] (Unrelated room and action)\n</summary>\n</think>\n<answer> [4] > [1] > [5] > [2] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_AR-VPPtV7ag.mp4", "activitynet/videos/v_jNGa0jPAMjI.mp4", "activitynet/videos/v_G9m5zxl2f08.mp4", "activitynet/videos/v_---9CpRcKoU.mp4", "activitynet/videos/v_Gd6YooP1_CU.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A person is holding a flute. The person shows a fingering for the second octave of the note. The person shows the fingering on the third octave. The woman talks directly to the camera.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a person assembling and holding a flute, with several frames focusing on fingering on the flute. The end frames show the person talking to the camera, but it is unclear what they are focusing on regarding octave fingerings.\n\n[2] The video includes detailed fingering charts with a focus on different flute octaves. The person clearly demonstrates fingerings on a flute corresponding to the notations shown, and the last row shows a woman talking directly to the camera.\n\n[3] A person holds and plays the flute, but no specific reference to octave fingerings is visible. The primary focus appears to be on playing rather than instruction or demonstration.\n\n[4] The video depicts a person playing the flute in a performance setting with no specific instructional content on flute fingerings or octave transitions.\n\n[5] The content is centered on a saxophone with instructional annotations. It is entirely unrelated to the flute or the specific query.\n</content>\n\n<contrast>\n- [2] beats [1] because it includes clear octave-specific notations and a demonstration of fingerings, while [1] lacks clarity on octave transitions.\n- [1] beats [4] because it includes some flute handling and interacting with the camera, whereas [4] focuses on performance without instructional elements.\n- [4] beats [3] because [4] shows clear flute performance, while [3] lacks clarity in its intent and relation to the query.\n- [3] beats [5] because while [3] involves a flute, [5] has no relation to flutes, focusing instead on a saxophone.\n- [3] and [4] both do not fulfill the query entirely but [4] has clearer flute context compared to [3].\n</contrast>\n\n<summary>\n1st=[2] (Includes specific fingering transitions and octave notations related to the query)\n2nd=[1] (Shows a person holding the flute and talking to the camera but unclear on octave transitions)\n3rd=[4] (Performance focus without direct relevance to the query specifics)\n4th=[3] (Shows flute but lacks clarity on fingering and octave information)\n5th=[5] (Focuses on saxophone, irrelevant to the flute query)\n</summary>\n</think>\n<answer> [2] > [1] > [4] > [3] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_P6ijX0vabJs.mp4", "activitynet/videos/v_p9iE8iemCMs.mp4", "activitynet/videos/v_f5xSylPY_bQ.mp4", "activitynet/videos/v_J6ScF5n_Cug.mp4", "activitynet/videos/v_oTKYejnHLtE.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Three girls are on the beach talking. They do several flips on the beach. They lift one of the girls up for a stunt.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The scene shows three girls on a beach. They perform exercises, flips, and a stunt where one girl is lifted by the others, matching the query well with beach location and activities described.\n\n[2] Contains various gym scenes with two girls but lacks the beach environment and specific stunts described in the query.\n\n[3] Shows a group of girls in a gym setting practicing lifts and stunts but not on a beach and without the full context of the query.\n\n[4] Consists of two girls talking and playing on the beach, but without any flips or stunts, partially matches the query.\n\n[5] Displays a cheerleading setting with group stunts and lifts similar to the query but in an indoor environment rather than a beach setting.\n</content>\n\n<contrast>\n- [1] ranks higher than [4] because it includes the beach setting and the specific activities of performing flips and stunts.\n- [3] beats [2] as it shows cheerleading elements like lifting and stunts, closer to the query.\n- [4] is slightly relevant because of the beach setting but lacks the stunt activities, ranking lower than [1].\n- [5] shows stunts and lifts in a cheerleading context, relevant to the query actions but not the location, ranking lower than [3].\n</contrast>\n\n<summary>\n1st=[1] (matches beach location, flips, and stunts)\n2nd=[3] (shows cheerleading stunts, but indoors)\n3rd=[5] (has cheerleading stunts, indoors)\n4th=[4] (beach setting, but no stunts)\n5th=[2] (indoor gym, no relevant stunts or beach)\n</summary>\n</think>\n<answer> [1] > [3] > [5] > [4] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_4x7HYKV8zkY.mp4", "activitynet/videos/v_w8LDUu66JWI.mp4", "activitynet/videos/v_II0BlR0BriI.mp4", "activitynet/videos/v_6-P9SRNbwac.mp4", "activitynet/videos/v_3Edxa-t4MxA.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"The camera pans above lush and green terrain over what looks like a mountain range. A man is shown covered in sweat and then jumping across various hills performing stunts on a dirt bike. He rotates and flips and performs all sorts of acrobatics. He is highly skilled and jumping over everything. The video then shows multiple men on dirt bikes jumping in various different places and shows the action from a variety of angles. The men continue performing their stuns in various other landscapes.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Displays various dirt bike stunts in outdoor environments. Scenes include flips and jumps across diverse landscapes. Shows multiple riders, fitting the query well.\n[2] Primarily features a young boy performing dirt bike maneuvers on simpler terrain. Focuses on close-ups of the child with few advanced stunts.\n[3] Includes professional shots of dirt bikes, focusing on motion and action, with clear branding. Shows stunt biking in a dedicated area but lacks diverse landscapes.\n[4] Features intense dirt bike tricks and flips, with riders performing acrobatics on dirt tracks. Scenes are dynamic and action-focused.\n[5] Shows a freestyle motocross event in an urban setting with large crowds. Highlights event atmosphere and complex aerial stunts.\n</content>\n\n<contrast>\n- [1] captures multiple riders and diverse landscapes, fitting the query better than [2], which focuses on a single young rider.\n- [1] and [4] both show advanced stunts, but [1] includes landscape variety.\n- [3] has professional captures but lacks landscape diversity, unlike [1].\n- [5] emphasizes an event atmosphere more than landscape and varied location stunts seen in [1].\n- [4] and [1] both focus on stunts, with [1] offering more landscape variety and contexts.\n</contrast>\n\n<summary>\n1st=[1] (Captures diverse landscapes, multiple riders, and acrobatic stunts)\n2nd=[4] (Focus on intense stunts, lacks varied landscapes)\n3rd=[5] (Focuses on an event atmosphere and complex stunts)\n4th=[3] (Professional captures, limited landscape variety)\n5th=[2] (Focus on a young rider with simpler maneuvers)\n</summary>\n</think>\n<answer> [1] > [4] > [5] > [3] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v__Af_9cK5x4E.mp4", "activitynet/videos/v_KCOzA0Suy5U.mp4", "activitynet/videos/v_QtCNHRtycmg.mp4", "activitynet/videos/v_p6jip892RdM.mp4", "activitynet/videos/v_v-dxQNxdMrU.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Athletes throw a javelin during a competition. A man practices his throwing stance walking and doing half throws without javeline. A man practices his throwing stance walking and doing half throws holding a javeline. The man swings a hammer. The man throws a weighted ball. Athletes train in a gym doing throwing motions on weighted cable machines.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video shows a javelin competition with athletes actively throwing javelins on a field. The sequence primarily focuses on javelin actions during a competitive event.\n\n[2] The video captures scenes of athletes practicing, likely in a discus or shot put cage. No javelin actions are visible.\n\n[3] This video features various training activities involving javelin stances, half throws, and resistance training with cable machines, aligning well with the query details.\n\n[4] Various athletes are shown in sequences of javelin throwing during competitions, consistent with the query's initial focus on javelin throwing.\n\n[5] The focus here is on shot put or discus events, with several athletes throwing weighted implements, which matches part of the query related to throwing weighted balls.\n</content>\n\n<contrast>\n- [3] beats [1] because it includes a broader variety of javelin-related training including both stance practice and cable machine training, whereas [1] focuses solely on competitive throws.\n- [4] is preferred over [5] because it contains more accurate representations of javelin throwing events, whereas [5] primarily shows shot put, less relevant to the complete query.\n- [1] is higher than [2] as it shows actual javelin use, whereas [2] shows other throwing disciplines without javelin involvement.\n- [2] beats [5] due to more visible athletic training, though neither fully matches the javelin query requirements.\n</contrast>\n\n<summary>\n1st=[3] (captures diverse practices including stance and cable training)\n2nd=[1] (focus on javelin competition with actual throws)\n3rd=[4] (includes competition throws, less training variety)\n4th=[2] (training focus, lacks javelin specifics)\n5th=[5] (focus on shot put, less relevant to javelin)\n</summary>\n</think>\n<answer> [3] > [1] > [4] > [2] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_GKK32IWDlSE.mp4", "activitynet/videos/v_uBhuULfs3Hc.mp4", "activitynet/videos/v_HxJnPUqF9i8.mp4", "activitynet/videos/v_aEUbl9oT_0s.mp4", "activitynet/videos/v_zLF5DSuDixg.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man is giving advice on outdoor activities, specifically mushrooming and closeup is shown and the man talks about it. The man cuts one open and then cuts it in half. Another man talks about the experience.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a man near a vehicle and a camping trailer, but there is no focus on mushrooming, cutting, or close-ups of mushrooms. Outdoor activities are present, yet they do not align closely with the query regarding mushrooming.\n\n[2] A man is discussing mushrooms, showcasing them in a close-up, cutting one open, and then in half, which perfectly matches the query description. Another individual is seen giving commentary, aligning it well with the requested experience sharing.\n\n[3] This features a man in a kitchen environment handling and peeling what appears to be a potato, not a mushroom. The setting is not outdoors, and the activity doesn’t match mushrooming specifics.\n\n[4] The clip reflects a discussion on soil and plant growth, not related to mushrooming or cutting mushrooms. The activity involves walking and inspecting soil, but no mushrooms are involved.\n\n[5] The content shows two individuals engaged in leather work inside a workshop, demonstrating polishing techniques, unrelated to outdoor activities or mushrooming. The context is entirely different from the query focus.\n</content>\n\n<contrast>\n- [2] features a man showcasing and cutting a mushroom, whereas [1] lacks any activity related to mushrooming or cutting mushrooms.\n- [2] is focused on mushrooms, including discussions and demonstrations, while [3] depicts indoor kitchen activity with vegetables, unrelated to the query.\n- [2] matches the focus on outdoor mushroom activities, contrasting with [4], which concerns plant soil inspection.\n- [2] involves mushroom cutting closely, unlike [5], which centers on a workshop task.\n</contrast>\n\n<summary>\n1st=[2] (Direct match: man giving advice on mushrooming, close-up, cutting)\n2nd=[1] (Outdoor activities present but no focus on mushrooming)\n3rd=[4] (Outdoor educational activity but not about mushrooming)\n4th=[3] (Indoor kitchen activity, unrelated to the query)\n5th=[5] (Indoor workshop focus, not connected to mushrooming)\n</summary>\n</think>\n<answer> [2] > [1] > [4] > [3] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_1LrEiYVd3qg.mp4", "activitynet/videos/v_UYHgieL5IKM.mp4", "activitynet/videos/v_rHWOESWciSc.mp4", "activitynet/videos/v_SIeV0DPSpHQ.mp4", "activitynet/videos/v_c-3ed17J3F8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A young child is seen sitting behind a drum set playing the drums. The boy continues playing the instrument while the camera zooms in on him playing.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] An adult is seen playing a drum set, with no clear indication of a child being present. The camera appears to zoom in on the drummer.\n[2] An adult drummer is actively playing drums. The sequence suggests a focus on his drumming, but no child is visible.\n[3] A young child is playing a drum set. The sequence clearly shows the child playing drums from different angles, indicating an emphasis on the child's action.\n[4] A young boy is shown playing a drum set. The sequence captures him drumming, with some variation in camera angle and zoom.\n[5] A young child is playing conga drums rather than a conventional drum set. The focus is on the child's interaction with the instrument.\n</content>\n\n<contrast>\n- [3] captures a young child playing a traditional drum set throughout, with clear focus and action visibility, meeting the query best.\n- [4] features a young boy actively playing a drum set, but the camera work is less dynamic than [3].\n- [1] shows an adult drummer, whereas [3] fits the query by showcasing a young child.\n- [5] features conga drums, not a drum set, whereas [3] accurately presents a traditional drum set.\n- [2] focuses on an adult drummer, lacking the child element present in [3].\n</contrast>\n\n<summary>\n1st=[3] (child playing drum set, clear and focused)\n2nd=[4] (young boy playing drum set, less dynamic angles)\n3rd=[1] (adult drummer, some zooming)\n4th=[5] (child with conga drums, not a drum set)\n5th=[2] (adult drummer, lacks child focus)\n</summary>\n</think>\n<answer> [3] > [4] > [1] > [5] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_90LkAH4tZC4.mp4", "activitynet/videos/v_ItFq_6cIByw.mp4", "activitynet/videos/v_nvsK9ggeics.mp4", "activitynet/videos/v_q3-S2tg4ULw.mp4", "activitynet/videos/v_FMUaey1tigI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"The video begins with a title sequence. A young man is shown in a gym performing tricks with a jump rope as music plays in the background. He then lays down and begins doing tricks on the floor. The young man gets up again and does more tricks while standing. At one point, his friends assist with the jump rope tricks. The video ends with credits showing the young man's name, along with the names of those who assisted him.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a gym setting where participants perform coordinated jump rope tricks. Multiple people assist in some tricks, showcasing teamwork but lacks a clear title or credits sequence.\n\n[2] This video opens with a title indicating \"Advanced Jump Rope Training\" and features an individual performing jump rope tricks in a studio environment. There is no sequence of floor tricks or presence of assisting friends.\n\n[3] The video starts with an animated title sequence, then focuses on a person performing breakdancing moves on the floor. There is no jump rope involved, nor are there visible assisting friends or jump rope tricks.\n\n[4] The video depicts an individual in a studio setting performing jump rope tricks, both while standing and on the floor. It ends with a credit frame, consistent with the query elements.\n\n[5] The video captures a public performance with someone doing jump rope tricks in a competitive or exhibition setting. There are no visible sequences illustrating a title or credits involving assisting friends.\n\n</content>\n\n<contrast>\n- [4] is ranked higher than [2] because it includes both standing and floor tricks, along with visible credits, aligning more closely with the query.\n- [2] features a title and jump rope tricks but lacks floor tricks and a credit sequence, ranking it below [4].\n- [1] shows group jump rope activities, but there's no clear individual focus or credits, placing it below [2].\n- [5] features jump rope tricks but lacks an indoor training/credit context, unlike [1].\n- [3] focuses on breakdancing moves, not jump rope tricks, ranking last as it significantly differs from the query.\n\n</contrast>\n\n<summary>\n1st=[4] (includes standing and floor tricks, ends with credits)\n2nd=[2] (title, jump rope tricks, no floor tricks or credits)\n3rd=[1] (group activity, no individual credit/title sequence)\n4th=[5] (public setting, lacks training/credit context)\n5th=[3] (no jump rope, focuses on breakdancing)\n</summary>\n</think>\n<answer> [4] > [2] > [1] > [5] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_ABB755sPZfY.mp4", "activitynet/videos/v_MlbM7Mew0Ys.mp4", "activitynet/videos/v_A0F6uT95xhA.mp4", "activitynet/videos/v_y7WO-8cQfO0.mp4", "activitynet/videos/v_Zhfm2dWtnpc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A lady is seen putting wax on a mans shoes. We see the cloth hanging on the wall. We see the footprints on the platform. The woman rubs the mans shoes with a cloth. We see the cloth on the wall, then the lines of cloth near the shoes shine sign. A man is standing on a platform. We see the holographic man on the wall with the paper.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Various scenes of shoe shining and polishing, but no cloth on the wall or holographic man visible. Focus on craftsmanship. \n[2] Showcases a man polishing shoes in a shop setting, but no woman, cloth on the wall, or holographic man present. \n[3] A woman is polishing a man's shoes with cloth; cloths and a holographic man are visible on the wall. Appears closely related to the query.\n[4] Shows a man cleaning sneakers in a room, not relevant to the shoe polishing context or query details.\n[5] A man polishing shoes on a platform, fitting some aspects of the query but lacks presence of a woman or holographic man.\n\n</content>\n<contrast>\n- [3] is ranked highest as it closely matches the query with a woman polishing shoes, cloths, and a holographic figure visible.\n- [5] ranks next by including shoe polishing on a platform, aligning partially with the query.\n- [2] is placed third as it shows shoe polishing but lacks other query elements.\n- [1] is fourth; it depicts shoe polishing but misses key elements like a woman or holographic man.\n- [4] is last; it involves sneaker cleaning unrelated to the query specifics.\n\n</contrast>\n<summary>\n1st=[3] (Close match: Woman polishing shoes, cloths and holographic man seen)\n2nd=[5] (Shoe polishing on a platform, partial match)\n3rd=[2] (Shoe polishing in general, but lacks query specifics)\n4th=[1] (Shoe polishing but misses key elements)\n5th=[4] (Sneaker cleaning, unrelated to query context)\n</summary>\n</think>\n<answer> [3] > [5] > [2] > [1] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_nEcOF04KK0g.mp4", "activitynet/videos/v_r34VM4v0ShI.mp4", "activitynet/videos/v_74EoyLUbtgI.mp4", "activitynet/videos/v_UsqhZb0co6o.mp4", "activitynet/videos/v_1OJa2iiFxfk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Two girls stand between a girl in a hair salon stool and begin brushing her and parting her hair. One woman braids her hair continuously to eventually create a french braid. Both women stand talking to the camera while presenting the braid just made.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A woman is seen alone styling her hair, braiding it herself while looking in a mirror. There is no indication of other people interacting with her hair as described in the query.\n[2] A woman is braiding another woman's hair in a salon setting. Two women are seen talking, but one is braiding continuously, matching the query's action of creating a French braid.\n[3] A woman is braiding her own hair, with no interaction from others. The setup is not in a professional salon environment and lacks the presence of two other girls.\n[4] A woman is braiding a younger girl’s hair in a domestic setting with another woman present, exhibiting parts of the query; however, it lacks a salon context and the presence of two helping parties.\n[5] A woman is braiding another woman's hair with a clear salon backdrop present. However, the focus is on the braiding process itself rather than having a second person assisting.\n</content>\n\n<contrast>\n- [2] matches the query because it features two women assisting in braiding hair in a salon setting, leading to a French braid.\n- [4] involves two people, but it misses the salon environment and dynamic interaction mentioned in the query, whereas [2] captures these elements.\n- [5] offers a salon setting with a braid but lacks the interaction with a second person as seen in [2].\n- [1] and [3] focus on individuals styling their own hair, with no visible assistance from two others, making [2] much more relevant.\n</contrast>\n\n<summary>\n1st=[2] (two people braiding in a salon setting, matching description)\n2nd=[4] (two people involved, lacks salon setting)\n3rd=[5] (salon setting with one person braiding)\n4th=[1] (solo hair braiding, no additional people)\n5th=[3] (solo hair braiding, no additional people)\n</summary>\n</think>\n<answer> [2] > [4] > [5] > [1] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_Tovbyt_ZiIc.mp4", "activitynet/videos/v_WY9zvUkK_4Y.mp4", "activitynet/videos/v_4b2_OpAGwW0.mp4", "activitynet/videos/v_cS19K2310Lc.mp4", "activitynet/videos/v_1FMjfLRpXwM.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man sits in a chair as he is getting his hair cut. A man in a black shirt moves a light around. People scramble behind the hair stylist. A hair stylist continuously combs and cuts the man's hair. The hair stylists rubs some pomade on the man's hair. The man is seen posing.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A man in a black shirt is styling a woman's hair, using a spray and comb. It focuses on the techniques used for hair styling, but not on a man getting his hair cut or people scrambling.\n\n[2] A man is having his beard and hairline trimmed, focusing on grooming rather than a full haircut or a photo shoot atmosphere.\n\n[3] A man is getting a haircut in a salon. The progression shows hair cutting and styling, but lacks any photo shoot activity or people moving around behind the stylist.\n\n[4] A stylist is giving a man a haircut with clippers and scissors. It involves styling but no visible photo shoot or extra people.\n\n[5] A man is having his hair cut and styled in a salon. The sequence includes a man posing at the end, with some activity suggesting a professional environment, aligning with the query about people moving around behind the stylist.\n</content>\n\n<contrast>\n- [5] better matches the query because it includes salon activity, hair cutting, and a man's pose, whereas [3] lacks posing or additional people involved.\n- [1] involves hair styling but focuses on a woman and lacks a photo shoot environment, whereas [4] has a haircut but no posing or photo shoot elements.\n- [3] and [4] both involve hair cutting, but [4] lacks the elements of a professional setup with people scrambling.\n- [2] mainly covers beard trimming, which does not match the full haircut and photo shoot context in [5].\n</contrast>\n\n<summary>\n1st=[5] (shows haircutting, styling, posing, and photographer activity)\n2nd=[4] (shows haircut and stylist activity, lacks photo shoot aspects)\n3rd=[3] (shows haircutting but lacks photo shooting and posing)\n4th=[1] (focuses on women's hair styling; no photo shoot environment)\n5th=[2] (focuses on beard grooming rather than haircutting context)\n</summary>\n</think>\n<answer> [5] > [4] > [3] > [1] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_7_5VT6QCqNI.mp4", "activitynet/videos/v_04LdesS7Pxk.mp4", "activitynet/videos/v_wr2FC2P3hIs.mp4", "activitynet/videos/v_g-Egl6eXlX4.mp4", "activitynet/videos/v_JgHQCKz_DRc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A group of people are on a field. They attach a camera to one of the player's heads. They lift each other over their heads in cheerleading stunts that can be seen from the camera's view. They smile and laugh as they continue to perform stunts.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This collage shows cheerleaders on a field performing individual stunts and exercises. There is no evidence of a camera being attached to a player's head or group stunts being captured from such a perspective.\n\n[2] This sequence occurs indoors and shows cheerleading routines with an audience. The video doesn't capture a field setting or suggest a camera attached to a player’s head for a unique perspective.\n\n[3] The images depict a black-and-white cheerleading performance. Stunts are visible, but the setting isn't clearly a field, and there's no indication of a head-mounted camera.\n\n[4] These frames showcase indoor cheerleading practice and performances. There is no sign of activities on a field or the unique camera perspective described in the query.\n\n[5] This sequence shows people outdoors on a field. A camera is visibly attached to one person's head, capturing stunts from their viewpoint. Participants are visibly smiling and laughing while performing.\n</content>\n\n<contrast>\n- [5] beats [1] because it features the specific action of a camera viewpoint attached to a player's head, whereas [1] shows individual stunts without that perspective.\n- [5] beats [2] because it takes place outdoors on a field, fitting the query better than [2]'s indoor performance.\n- [5] beats [3] as it shows smiling and laughing participants along with the camera perspective, whereas [3] lacks the specified camera viewpoint and outdoor setting.\n- [5] beats [4] due to the outdoor field setting and the use of a head-mounted camera, both absent in [4].\n- [3] beats [4] because, although [3] lacks specific details, it maintains closer relevance with cheerleading stunts visible.\n- [2] beats [1] by showing group stunts and having an audience which suggests an event-based context, closer to the query's collaborative nature.\n</contrast>\n\n<summary>\n1st=[5] (features relevant camera perspective and field location)\n2nd=[3] (shows stunts despite missing camera perspective and field)\n3rd=[4] (contains cheerleading indoors without correct setting)\n4th=[2] (indoor audience event doesn't match field setting)\n5th=[1] (individual stunts miss camera angle and group collaboration)\n</summary>\n</think>\n<answer> [5] > [3] > [4] > [2] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_DbF6oI705UA.mp4", "activitynet/videos/v_vPhAXENSjiw.mp4", "activitynet/videos/v_BSl22Hx2WGM.mp4", "activitynet/videos/v_d0h_l9bE1o0.mp4", "activitynet/videos/v_Vjsd6Ec6eKA.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Music plays as surfers ride the waves. A few men are shown preparing to surf, prepping their board and getting dressed. A helicopter rides over the shoreline where the surfers are riding. A couple laughs together on the beach in a tent awaiting more surfing. A man is then shown prepping his board with surf wax. The crowd watches others riding waves. The crowd cheers multiples surfing runs on. Kelly Slater is featured on a successful run, followed by more clips riding the waves. Kelly Slater is paused mid wave to wrap up the story line. Music and producer credits are then shown.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows various surfing scenes, including surfers riding waves, preparing their boards, and a crowd on the beach. Matches the surfing theme but lacks helicopters and specific references to Kelly Slater or a storyline wrap-up.\n[2] Features an individual surfing with close-up shots, possibly identifying Kelly Slater. Focuses on individual surfing maneuvers and facial close-ups, relevant but limited in broader context.\n[3] Displays dynamic surfing, large crowds watching, and competitive runs. Likely includes Kelly Slater's performance, fitting parts of the query well, including surfing highlights and crowd reactions.\n[4] Centers on a surfing competition with awards and achievements. Less about the surfing process and story but includes crowd and competition aspects.\n[5] Involves a separate extreme sport (skyboarding), diverging from the surfing theme. Lacks relevance to surfing or associated storyline elements.\n\n</content>\n\n<contrast>\n- [3] beats [1] because it clearly features surfing competitions with crowds and possibly Kelly Slater, matching the query’s highlights.\n- [1] beats [2] as it shows more diverse beach activities and crowd presence, slightly better matching the query.\n- [2] beats [4] as it focuses more on surfing rather than competition aftermath.\n- [4] beats [5] because it relates to surfing competitions, unlike the wakeboarding sequence in [5].\n- [5] is least relevant because it depicts skyboarding rather than surfing.\n\n</contrast>\n\n<summary>\n1st=[3] (Features surfing highlights, crowd, and possibly Kelly Slater)\n2nd=[1] (Includes surfing and prep activities, lacks specific features)\n3rd=[2] (Focuses on individual surfer, possible Kelly Slater, lacks full context)\n4th=[4] (Competition focus, less on surfing action)\n5th=[5] (Unrelated extreme sport content)\n\n</summary>\n</think>\n<answer> [3] > [1] > [2] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_ANB_IQfi9kA.mp4", "activitynet/videos/v_O9crPOB_9tE.mp4", "activitynet/videos/v_QdmUojrWdmE.mp4", "activitynet/videos/v_r0P0egQt-jU.mp4", "activitynet/videos/v_DHWmzuAWxLM.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A girl is seen talking to the camera with her hand on her hip and leads into her holding up various ingredients. She then mixes the ingredients together on a slice of bread and creates a sandwich while pouring chips out onto the side. She then presents the plate while speaking to the camera and holding up various liquids.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This sequence features a person in a uniform preparing a sandwich at a commercial food counter. The person is not holding the items or talking to the camera in a personal kitchen setting. It doesn't match the query involving a personal presentation with ingredients and sandwich-making at home.\n\n[2] This sequence shows a woman in a home kitchen setting talking to the camera with her hand on her hip, holding various ingredients, making and presenting a sandwich, and pouring out a drink. This perfectly matches the query with comparable actions and setting.\n\n[3] The sequence shows a woman in a kitchen setting presenting pre-arranged sandwich ingredients and talking to the camera, but does not show her holding ingredients or mixing them on bread in the way described in the query.\n\n[4] The sequence features a woman explaining sandwich ingredients and displaying them, then making and presenting a sandwich. However, there is less emphasis on speaking while making the sandwich compared to [2].\n\n[5] This sequence shows a woman preparing a grilled cheese sandwich, including cutting, grilling, and presenting. It doesn't fully align with the query as it focuses on grilling rather than a varied ingredient sandwich setup.\n</content>\n\n<contrast>\n- [2] is better than [4] because it matches the scenario of talking to the camera and preparing a sandwich with more ingredient interaction.\n- [4] is better than [3] because it shows direct sandwich assembly and presentation, not just explaining the ingredients.\n- [3] is better than [5] as it involves speaking while presenting sandwich ingredients, whereas [5] focuses mainly on grilled cheese preparation.\n- [5] is more relevant than [1] because it shows sandwich-making, while [1] is more commercial and less personal, not matching the home setting in the query.\n</contrast>\n\n<summary>\n1st=[2] (matches query with speaking, ingredient display, and sandwich making)\n2nd=[4] (similar actions but less emphasis on interaction)\n3rd=[3] (presents sandwich but less interactive)\n4th=[5] (focus on grilled cheese not varied ingredients)\n5th=[1] (commercial setting, less matching to query)\n</summary>\n</think>\n<answer> [2] > [4] > [3] > [5] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_78UikTQkGgg.mp4", "activitynet/videos/v_KE2tDW9x9d8.mp4", "activitynet/videos/v_LlI1IJEkW5E.mp4", "activitynet/videos/v_GtYf9VsPHO0.mp4", "activitynet/videos/v_GaxyzK2mHqw.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A boy surfs in a big wave. A group of people lye down on surfboards in a wave. Suddenly, a boy stands an surf below the wave. Men stand in the beach holding surfboards.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows various surfing scenes: a boy on a big wave, a group lying on surfboards, a boy surfing in a tube, and men standing with surfboards on the beach.\n[2] This video depicts surfers catching big waves, a boy effectively surfing under a wave, and men walking with surfboards on the beach.\n[3] The sequence involves surfing on big waves and some scenic views but lacks the specific actions described in the query.\n[4] There are surfing activities with multiple surfers in the water, but it does not fully align with the query’s elements.\n[5] Focuses on individual surfing and surfers on the beach but lacks the group lying on surfboards.\n</content>\n\n<contrast>\n- [2] is ranked higher than [1] because it clearly shows people surfing large waves, a boy surfing under a wave, and men with surfboards on the beach, matching all query elements.\n- [1] contains all elements but has more varied scenes that are less focused.\n- [5] excels in individual surfing shots and beach scenes but lacks the group lying down.\n- [4] has surfing and some group water scenes but lacks clear alignment with the boy surfing under the wave.\n- [3] focuses more on scenic and surfing footage but misses specific actions relevant to the query.\n</contrast>\n\n<summary>\n1st=[2] (Clear depiction of all query elements: big wave surfing, boy under wave, men on the beach.)\n2nd=[1] (Includes all query elements but with less focus and coherence.)\n3rd=[5] (Good individual surfing and beach sequences but lacking group actions.)\n4th=[4] (Some group surfing but misses key actions like boy under wave.)\n5th=[3] (Scenic and general surfing but lacks detailed action matching query.)\n</summary>\n</think>\n<answer> [2] > [1] > [5] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_O0uVjGlOW8M.mp4", "activitynet/videos/v_B3ZVUyLeT84.mp4", "activitynet/videos/v_ThRbaI4pGIs.mp4", "activitynet/videos/v_yrf93aLQXBE.mp4", "activitynet/videos/v_sVeC5VlDsjk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A girl with several facial piercings his seen rubbing lotion all over her face and powdering her cheeks and face. She then puts eyeliner on followed by mascara and lip liner. She colors in her eyebrows and puts on fake eye lashes, then rubbing some of the black off and more mascara on. She puts a bit of blush in her cheeks and then smiles and waves to the camera.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A person shows how to clean and insert contact lenses, applying some steps but not involving makeup. \n[2] A girl with piercings applies makeup, including lotion, powder, eyeliner, mascara, lip liner, filling eyebrows, and putting on fake eyelashes, closely matching the query.\n[3] Focuses on applying lip liner and possibly some makeup, but lacks the complete face preparation in sequence as per the query.\n[4] Shows a girl applying makeup on others, not on herself, and without facial piercings.\n[5] Displays information about cosmetics safety, not a makeup application process.\n\n</content>\n\n<contrast>\n- [2] beats [3] because [2] covers the entire makeup sequence including facial piercings, while [3] focuses more narrowly on lips.\n- [3] beats [1] because [3] involves face makeup application, whereas [1] involves contact lens instruction.\n- [1] beats [4] because [1] involves an individual procedure, while [4] involves applying makeup on others.\n- [4] beats [5] because [4] involves some makeup application, while [5] is informational on cosmetics.\n</contrast>\n\n<summary>\n1st=[2] (Comprehensive makeup application sequence with facial piercings, directly matching the query)\n2nd=[3] (Partial makeup application, lacks completeness and facial piercings)\n3rd=[1] (Contact lens tutorial, minor relevance but individual-focused)\n4th=[4] (Application on others, less relevant)\n5th=[5] (Informative content on cosmetics, not on application process)\n</summary>\n</think>\n<answer> [2] > [3] > [1] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_4WNso9fVJG8.mp4", "activitynet/videos/v_QeVVKZhoFCg.mp4", "activitynet/videos/v_avGe67wRvxA.mp4", "activitynet/videos/v_3ddoLUFrGsE.mp4", "activitynet/videos/v_RA61jh8HwPI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman is standing by a toilet in a bikini. A man is rubbing sunscreen onto her legs. She begins to rub her leg.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows people outdoors playing pool; no relation to the query.\n[2] Features a woman in a bathroom wrapping a towel and preparing to shave her legs; not relevant to query specifics.\n[3] Displays a woman on the beach applying sunscreen, but no man is rubbing it on her; partially relevant.\n[4] Shows two women in a bath setting possibly shaving or applying cream; not closely related to query details.\n[5] Shows a woman in a bikini standing by a toilet, with a man applying and rubbing sunscreen on her legs, followed by her rubbing her leg; directly matches the query.\n</content>\n\n<contrast>\n- [5] includes a woman in a bikini near a toilet and a man rubbing sunscreen on her legs, matching the query, whereas others do not directly fit the scenario.\n- [3] has sunscreen application but lacks the setting and the specific interactions, whereas [5] meets all criteria.\n- [2] and [4] involve leg-related activities but not sunscreen or the specific scene setting, unlike [5].\n- [1] is unrelated, focusing on pool play and lacking any elements of the query.\n</contrast>\n\n<summary>\n1st=[5] (directly matches query criteria)\n2nd=[3] (partial match with sunscreen application)\n3rd=[2] (has leg activity, but unrelated setting)\n4th=[4] (some leg activity, incorrect context)\n5th=[1] (unrelated content)\n</summary>\n</think>\n<answer> [5] > [3] > [2] > [4] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_6h-WE-0eBlA.mp4", "activitynet/videos/v_OoyoTLkSOuY.mp4", "activitynet/videos/v_MY6o5ZObFLE.mp4", "activitynet/videos/v_HeHQxObROwQ.mp4", "activitynet/videos/v_wBhtt9x74TA.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman is seen speaking to the camera while holding up various objects. The puts tape on he eye and begins putting eyeliner above her eye. She continues putting the makeup on while speaking to the camera and takes off the tape in the end.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a woman discussing and demonstrating the use of colored contact lenses. There is no evidence of her applying makeup with tape or eyeliner. The main focus is on the contact lenses.\n\n[2] The woman is seen in a bedroom applying various types of makeup to her face but does not use tape on her eye nor applies eyeliner. The focus is on makeup, but not in the specific way described.\n\n[3] The sequence features a woman discussing makeup, showing an eyeliner pencil, and applying it. There's no tape involved, and while the close-up eyeliner application is relevant, the use of tape is absent.\n\n[4] This video most directly matches the query, showing a woman speaking to the camera, applying tape near her eye, and then using eyeliner. The tape is visibly removed, fitting the entire sequence specified.\n\n[5] A woman is discussing and demonstrating contact lenses. There's no application of makeup or use of tape in this sequence. The focus is on handling contact lenses.\n</content>\n\n<contrast>\n- [4] is ranked highest because it clearly shows the use of tape and eyeliner, directly matching the query specifics, whereas [3] only shows eyeliner application without tape.\n- [3] comes second with eyeliner application, but lacks the tape aspect, whereas [2] involves make-up but not as specifically described.\n- [2] includes broader makeup application but lacks the query's specific tape and eyeliner actions, whereas [1] focuses on contact lenses instead of makeup.\n- [1] is more about contact lenses, making it less relevant than [5], which also centers on lenses but still more relevant than no direct makeup application at all.\n- [5] discusses contact lenses but is the least related to the makeup context of the query.\n</contrast>\n\n<summary>\n1st=[4] (Directly matches with tape and eyeliner application as described)\n2nd=[3] (Includes eyeliner application but lacks tape)\n3rd=[2] (Features makeup application but not specifically tape and eyeliner)\n4th=[1] (Focuses on contact lenses and unrelated to makeup specifics)\n5th=[5] (Contact lenses discussed, least related to makeup)\n</summary>\n</think>\n<answer> [4] > [3] > [2] > [1] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_Ie7DREiXp08.mp4", "activitynet/videos/v_aSPPD0fDI3g.mp4", "activitynet/videos/v_InA8Xbg-hvo.mp4", "activitynet/videos/v_f-aGlEjgZoE.mp4", "activitynet/videos/v_extd66Y1tJ8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man is seen standing in a field with a dog while holding onto toys. The man then begins throwing the toys while the dog chases after them. The men continues to do tricks with the dog and end with a picture.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a man with a dog indoors, performing tricks with props like discs. The dog retrieves and interacts with the man, demonstrating coordination but not in a field.\n[2] This video matches the query well, featuring a man outdoors in a field with a dog. The man throws discs, toys are visible, and the dog chases after them, culminating in play activities and interaction.\n[3] A long-range field view with what seems to be a competition. There is a field and a dog, but the detail level is low; actions aren't clear and interaction is not visible.\n[4] A man interacting with a dog in a field-like setting, visible throwing and retrieving actions. However, the scene is at night and doesn't fit the complete description ending with a picture.\n[5] Features a man in an open field with expansive views. The man performs tricks and throws with the dog, interacting with toys like discs. However, there’s no clear indication of taking a final picture.\n</content>\n\n<contrast>\n- [2] is better than [5] because it clearly shows toys, throwing, and chasing in a field, aligning with all query aspects, whereas [5] lacks the picture-taking.\n- [5] beats [1] because it's outdoors and closer to the described environment, whereas [1] is indoors.\n- [1] ranks higher than [4] due to clear, visible tricks and interactions even indoors, whereas [4] lacks ending interaction elements.\n- [4] is better than [3] because despite being nighttime, it has visible actions, unlike [3]'s wide view with low detail.\n- [3] ranks last due to no visible interaction or clear action fitting the query.\n</contrast>\n\n<summary>\n1st=[2] (matches all aspects of the query closely)\n2nd=[5] (close match but without final picture)\n3rd=[1] (indoor setting with visible tricks)\n4th=[4] (visible actions but no conclusion)\n5th=[3] (low interaction detail)\n</summary>\n</think>\n<answer> [2] > [5] > [1] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_lMbDEY-CVwE.mp4", "activitynet/videos/v_UhB739uudDo.mp4", "activitynet/videos/v_eLeGJBfqURA.mp4", "activitynet/videos/v_KoeytdNfQ_M.mp4", "activitynet/videos/v_Yd98o-oEGvI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A group of men are in a kayak going down the river bank. The are paddling and looking at the trees in the distance enjoying their day. One of the men gets out of the boat to go swimming in the water, They all get in and start splashing water, swimming under and enjoying the day before getting back in to turn back. Once they start to head back to land and to get ready to get on the dock.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows rafting on a river with a focus on paddling. There are brief moments of relaxation but no swimming or splashing. The activity ends with paddles held together, possibly symbolizing the end of the trip. \n\n[2] Features a group navigating rapids in an inflatable boat. Some individuals are in the water, which suggests swimming. However, the focus is primarily on maneuvering through the rapids, not leisurely enjoyment or returning to shore.\n\n[3] Depicts a leisurely canoe trip with two men paddling and swimming. It includes scenes of underwater swimming and enjoyment in the water, aligning well with the query about paddling, swimming, and enjoying the day.\n\n[4] Involves a canoe trip with moments of paddling and dock-related activities, but no evident swimming or water splashing, making it less relevant to the query.\n\n[5] Shows a single person kayaking along a calm river. It highlights paddling through peaceful waters but lacks any swimming or splashing action, focusing instead on a solo experience.\n</content>\n\n<contrast>\n- [3] ranks higher than [1] because [3] includes both paddling and swimming, whereas [1] focuses more on rafting with no swimming.\n- [3] beats [2] as it aligns better with the leisure and swimming elements of the query, whereas [2] is more focused on navigating rapids.\n- [1] is better than [5] due to group activity and relaxed moments, whereas [5] is a solo kayaking experience without swimming or group dynamics.\n- [1] also ranks higher than [4] as it involves more action and group interaction, whereas [4] lacks the swimming and group play elements.\n- [5] is placed above [4] as it involves kayaking, which is closer to the query than fishing, which is shown in [4].\n</contrast>\n\n<summary>\n1st=[3] (includes paddling, swimming, leisure, fitting the query well)\n2nd=[1] (group rafting with paddling, moderate relevance)\n3rd=[2] (features water activity but more focused on rapids)\n4th=[5] (solo kayaking, lacks group interaction or swimming)\n5th=[4] (canoeing with no swimming or play, least relevant)\n</summary>\n</think>\n<answer> [3] > [1] > [2] > [5] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_F_sbhegCsyg.mp4", "activitynet/videos/v_AXA-G5DdNrU.mp4", "activitynet/videos/v_djgM0IWIBKc.mp4", "activitynet/videos/v_KxWvKN_kGMY.mp4", "activitynet/videos/v_i69NhEDno_M.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man wearing a black clothes is sharpening the knife on a stone. The man beside him on blue long sleeves drawing something on the table with his finger. The man turn the knife to sharpen the other side and then wipe it with paper towel. The man in navy blue shirt point to the knife while the man sharpening the knife tries to sharpen it one hand. He continues sharpening the knife, turn it again to further sharpen the other side and wipe it again with paper towel. Throws the old and dirty paper towel and reach the roll of paper towel and clean the knife.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Two people are involved in sharpening. One uses a sharpening stone while another gestures nearby. No paper towel interaction or wiping action visible.\n[2] A man in a blue long-sleeve shirt sharpens a knife and wipes it using a cloth. Shows temporally coherent sharpening and cleaning actions.\n[3] A mechanical knife sharpener is used. Lacks relevance to manual sharpening and doesn't involve two people.\n[4] A person sharpens a knife and uses paper to test its sharpness. Focuses on sharpening but lacks clear interaction between two people.\n[5] Shows a person in black clothes sharpening a knife with a stone and cleaning it with a paper towel. Multiple people are around, possibly indicating interactions.\n</content>\n\n<contrast>\n- [5] beats [2] because it involves a man in black clothes performing actions more aligned with the query’s description, including multiple people and knife cleaning.\n- [2] beats [4] because it shows clear knife sharpening by a man in blue, closely aligning with the query but lacks a second clear person and some specific actions.\n- [4] beats [1] because it shows closer alignment with the query involving knife sharpening and cleaning, even though person interaction is minimal.\n- [1] beats [3] because it involves manual sharpening with partial relevance to some elements of the query while [3] is about mechanical sharpening.\n</contrast>\n\n<summary>\n1st=[5] (Man in black performs relevant actions, multiple people present)\n2nd=[2] (Man in blue sharpens knife, wipes it, close alignment)\n3rd=[4] (Knife sharpening and testing visible, lacks dual interaction)\n4th=[1] (Manual sharpening present, lacks coherence and specific actions)\n5th=[3] (Mechanical process misaligns with manual query)\n</summary>\n</think>\n<answer> [5] > [2] > [4] > [1] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_Anvhv9FFv1Q.mp4", "activitynet/videos/v_dtvIQ9Pd-Bg.mp4", "activitynet/videos/v_WLpjci5dN8s.mp4", "activitynet/videos/v_UzdI8PLRiXg.mp4", "activitynet/videos/v_E3h0UHRsSQk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"An intro leads into a woman pulling out a bucket and piecing together a mop. The camera pans around the bucket and the woman continues using the mop inside the bucket and moving around the object.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video begins with an introduction featuring a logo, followed by a woman assembling a mop beside a bucket. She uses the mop inside the bucket and moves it around, showcasing its use in a room setting.\n\n[2] A woman is outdoors discussing or performing gardening tasks. There is no mop involved, and it does not match the query.\n\n[3] The video showcases a woman using a window cleaning kit indoors, focusing on glass surfaces. It does not include a bucket or a mop setup.\n\n[4] The video includes scenes of a person mopping a floor, showing a bucket briefly. However, it lacks the detailed setup and indoor context specified in the query.\n\n[5] This video involves window cleaning tasks, with no relation to mopping or bucket assembly inside.\n</content>\n\n<contrast>\n- [1] is directly relevant because it shows the woman assembling a mop and using it with a bucket, as described.\n- [4] has some mopping action but lacks the specific assembly and indoor context described.\n- [3] features cleaning, but it is focused on windows, not mopping with a bucket.\n- [5] and [2] show unrelated tasks, window cleaning and gardening, respectively.\n</contrast>\n\n<summary>\n1st=[1] (directly matches query with mop assembly and use)\n2nd=[4] (includes mopping, but lacks indoor assembly context)\n3rd=[3] (cleaning content but unrelated to mopping or bucket)\n4th=[5] (features window cleaning, not mopping or indoors)\n5th=[2] (gardening content, unrelated to query)\n</summary>\n</think>\n<answer> [1] > [4] > [3] > [5] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_MSjSXzk2k50.mp4", "activitynet/videos/v_B8WIh6PUjE0.mp4", "activitynet/videos/v_E0DbrJVJUho.mp4", "activitynet/videos/v_IkXuooaGnCc.mp4", "activitynet/videos/v_Kq115s7yL3I.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A close up of a wooden bench is seen followed by a person mixing ingredients together. A person is then seen holding a rag and wiping down the bench. More ingredients are mixed afterwards that leads into more wiping down with the rag.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video guides viewers on making natural furniture polish. It starts with a close-up of a wooden surface, mixing olive oil and vinegar, and shows applying the mixture and wiping the bench.\n\n[2] This video demonstrates applying linseed oil on a wooden bench. It includes close-ups of mixing ingredients and multiple sequences of applying and wiping the bench.\n\n[3] The sequence depicts a person working on a wooden surface but focuses on applying a product rather than mixing ingredients. The action is consistent but lacks the repeated mixing and wiping seen in the query.\n\n[4] This video features various indoor scenes with a brief focus on applying products to a wooden surface, but doesn't emphasize mixing ingredients or repeated wiping.\n\n[5] The sequence shows a person in a workshop applying a product to wood. It involves less focus on mixing ingredients and repeated wiping, deviating from the query.\n</content>\n\n<contrast>\n- [2] beats [1] because it involves more sequences of mixing ingredients and wiping, closely matching the query.\n- [1] beats [3] because it includes visible mixing and wiping actions, fitting the query.\n- [3] beats [5] because it has more visible actions related to wood treatment, even if not mixing.\n- [5] beats [4] because of more consistent wood working actions and close-up views.\n- [4] ranks lower due to less relevance to mixing and wiping actions.\n</contrast>\n\n<summary>\n1st=[2] (mixing and wiping sequence matches query)\n2nd=[1] (clear mixing and wiping, though less extensive)\n3rd=[3] (focus on wood application, less on mixing)\n4th=[5] (consistent application, but less overall mixing)\n5th=[4] (less relevant to query specifics)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [5] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_1ATBCBUgflA.mp4", "activitynet/videos/v_MiTIi_QYt2k.mp4", "activitynet/videos/v_yOKDaH8Go0A.mp4", "activitynet/videos/v_pn1XDTlNMso.mp4", "activitynet/videos/v_0fw8it7Gj7k.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Men play badminton in the sand. A man on the left serves the birdie both men on the right miss it. the man on the right serves and the man on the left hits it back but it hits the net. A man stands on the left watching.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows men playing volleyball on a sandy beach and not badminton. Involves net play and multiple people, no indication of rackets or birdies.\n[2] Takes place indoors on a wooden court, involves racquet sports but not on sand. Doesn't match the sand setting.\n[3] Men are playing volleyball on the sand. There is no indication of badminton activity, similar lack of rackets and birdies.\n[4] Clearly displays two men playing badminton on a sandy court. Actions similar to query: serving and hitting the net.\n[5] Features a sand volleyball match with a noticeable crowd and scoreboard, no badminton gear evident.\n</content>\n\n<contrast>\n- [4] matches the badminton and sand context, whereas [1], [3], and [5] show volleyball.\n- [2] is played indoors, whereas [4] is appropriately in sandy outdoor setting.\n- [4] displays badminton rackets and a net, whereas others lack rackets.\n</contrast>\n\n<summary>\n1st=[4] (matches badminton on sand with rackets)\n2nd=[1] (beach setting aligns partially, but it's volleyball)\n3rd=[3] (beach setting with action, but it's volleyball)\n4th=[5] (beach sand setting, volleyball not badminton)\n5th=[2] (indoor play does not match sand context)\n</summary>\n</think>\n<answer> [4] > [1] > [3] > [5] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_MO8Dfz2Lp3w.mp4", "activitynet/videos/v_vlwOSDuiKQA.mp4", "activitynet/videos/v_TnYCazY5WEM.mp4", "activitynet/videos/v_95pMelnv2dQ.mp4", "activitynet/videos/v_ZWHFbIjh0N8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A class of kids are at outside at school dressed in uniforms with their teacher standing next to them. As they are in line,one girl is in the middle of a game of hop scotch,turns around,finishes and gets back in line. Once she is finishes,all of the other kids takes their turn one by one.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a group of kids in school uniforms playing hopscotch outside. A teacher is present, and one girl is playing while others are in line to take turns, consistent with the query.\n\n[2] Features a couple of children playing hopscotch on a sidewalk, but not in a school setting with uniforms or a teacher present.\n\n[3] Displays children in a school environment with uniforms, but they're not playing hopscotch. The focus is on general outdoor activities.\n\n[4] Shows a girl playing on a hopscotch-like grid, but not in a school or group setting with uniforms and a teacher.\n\n[5] Depicts a shuffleboard game, not related to hopscotch or the described school scenario.\n</content>\n\n<contrast>\n- [1] is a better match because it closely depicts the school setting, uniforms, hopscotch, and teacher presence, whereas [2] lacks this specific school context and uniforms.\n- [1] shows the exact sequence described with children in line and a teacher, while [3] does not involve hopscotch or sequence of turns in a game.\n- [1] involves multiple children and a structured activity in a school setting, unlike [4], which shows an individual child without school context or uniforms.\n- [1] involves hopscotch, matching the query, whereas [5] shows a completely unrelated game.\n</contrast>\n\n<summary>\n1st=[1] (accurately follows the query with school setting, uniforms, hopscotch, and teacher present)\n2nd=[2] (features hopscotch but lacks school context and uniforms)\n3rd=[3] (shows school setting and uniforms but not hopscotch or specific game sequence)\n4th=[4] (shows hopscotch-like activity but not in a school setting)\n5th=[5] (unrelated to hopscotch or school setting)\n</summary>\n</think>\n<answer> [1] > [2] > [3] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_Fk0z6MeCV0I.mp4", "activitynet/videos/v_l8NepxsCh34.mp4", "activitynet/videos/v_dZ4y7J3ASEE.mp4", "activitynet/videos/v_pieT6K771WQ.mp4", "activitynet/videos/v_yW76UI8lnNc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A large group of cheerleaders are seen running into an arena cheering and beginning performing a routine with one another. The girls continue cheering around the stage while lifting one another up and performing tricks and pans to the audience in the end.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This sequence depicts a group of cheerleaders running into an arena, performing stunts, and interacting with a large audience. The actions include cheerleading routines with lifts and coordinated movements, with visible panning towards the audience at the end.\n[2] Shows a group practicing cheerleading or gymnastics indoors. There are lifts and routines, but no evidence of a large audience or arena setting.\n[3] Displays a series of cheerleading events in a competitive setting, but lacks a clear narrative of entering an arena and does not show a continuous routine or lifting sequence with audience interaction as requested in the query.\n[4] Displays cheerleaders performing in a gym with an audience, featuring coordinated movements and formations but less emphasis on entering the arena or lifting sequences.\n[5] Shows cheerleaders performing a routine with audience present. There are coordinated movements and lifting, but the entry into the arena is not as emphasized or clear.\n</content>\n\n<contrast>\n- [1] best matches the query with a clear sequence of entering, performing routines, lifting, and panning to an audience.\n- [5] whereas highlights a large audience and performance but lacks an arena entry focus.\n- [4] focuses on performance and audience interaction but less on entry or lifting coherence.\n- [2] focuses on practice rather than an arena setting, lacking audience interaction.\n- [3] whereas provides competitive glimpses without entry or clear routine detailing.\n</contrast>\n\n<summary>\n1st=[1] (enters arena, performs routines, lifts, audience interaction)\n2nd=[5] (audience present, routines, some lifting)\n3rd=[4] (performance, audience, less lifting focus)\n4th=[2] (practice setting, lacks arena/audience)\n5th=[3] (competition clips, less narrative coherence)\n</summary>\n</think>\n<answer> [1] > [5] > [4] > [2] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_O4P07fipvIA.mp4", "activitynet/videos/v_DEmArEKL8gc.mp4", "activitynet/videos/v_mRe2xY_amWw.mp4", "activitynet/videos/v_fZ_CmvF-_No.mp4", "activitynet/videos/v_opAcl-kukdU.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Two kids are climbing a rock wall. The adults are guiding the lines. The kids continue to climb. The one on the right will make it to the top first. The one on the left did not make it to the top, both hop back down the wall.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows two kids climbing a rock wall with adults guiding the lines. The kids are climbing side by side, and the one on the right is ahead of the one on the left. Eventually, the right climber reaches the top first, while the left climber does not make it to the top. Both climbers then descend the wall.\n[2] This video involves one person explaining and demonstrating climbing techniques on an indoor climbing wall. There are different views of the wall, but no children or clear sequences matching the query's detailed climbing competition.\n[3] This video depicts children riding camels in an outdoor setting, with no evidence of rock climbing or related activities.\n[4] The video shows individuals climbing a large rock wall over water with nighttime conditions, but does not feature children, adults guiding them, or the specific climbing sequence described.\n[5] This video involves a man explaining climbing equipment and techniques on an indoor wall. It lacks children climbing or adult guidance, and does not match the specific climbing scenario in the query.\n</content>\n\n<contrast>\n- [1] is highly relevant as it shows two kids climbing a rock wall with adults guiding them, matching the query sequence and outcome, whereas [2] is instructional and not focused on kids climbing.\n- [1] beats [4] as [4] shows a night-time climbing event over water with adults, not children with guided climbs as in the query.\n- [1] beats [5] due to its focus on kids climbing with guidance, whereas [5] is centered around gear and technique instruction without the described action sequence.\n- [3] is a mismatch entirely, showing camel rides, thus least relevant.\n</contrast>\n\n<summary>\n1st=[1] (kids climbing a rock wall with adult guidance and the correct sequence)\n2nd=[2] (indoor climbing demonstration, lacks kids or specific sequence)\n3rd=[5] (gear and technique instruction, no kids climbing)\n4th=[4] (night climbing event over water, no kids or sequence)\n5th=[3] (camel rides, no climbing activity)\n</summary>\n</think>\n<answer> [1] > [2] > [5] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_XHMrvd3sKzE.mp4", "activitynet/videos/v_jVM8v6uJx8c.mp4", "activitynet/videos/v_s9Wop4PMZaw.mp4", "activitynet/videos/v_etfSrbmOvUE.mp4", "activitynet/videos/v_q4rVY3sLQqA.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man walks up to a snow scoop. The man shovels the snow on a roof while the cameraman talks. The man stops scooping the snow.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a person shoveling snow near a house, but no evidence of a roof or a talking cameraman is visible. The activity seems related to shoveling on the ground.\n\n[2] A man actively uses a snow scoop on a snowy surface presumably on the ground. The man works with the scoop throughout, engaging in snow clearing but no visible roof or talking cameraman.\n\n[3] This video shows a woman at night clearing snow from a car, not a roof. The setting is different from the query requirements, lacking a snow scoop or a talking cameraman.\n\n[4] Features a man initially talking directly to the camera before shoveling snow on the ground. Although there is speech, the scene doesn’t involve a roof or snow scoop.\n\n[5] Shows a man working on a roof, but the task involves applying roofing materials, not shoveling snow. This video does not match the query as there is no snow involved.\n</content>\n\n<contrast>\n- [2] shows the most relevant snow clearing activity using a snow scoop, whereas [1] involves shoveling but not on a roof.\n- [1] involves shoveling snow, which is more relevant than [4], where the man talks and shovels without reaching a roof scene.\n- [4] includes speaking, but the snow shoveling is on the ground, unlike [5] which involves no snow-related tasks.\n- [5] involves roof work but not snow shoveling, unlike [3], which features a different task entirely.\n- [3] shows nighttime snow clearing from a car, irrelevant to the query’s focus on a roof.\n</contrast>\n\n<summary>\n1st=[2] (man uses snow scoop throughout consistently, most relevant task execution)\n2nd=[1] (engages in snow shoveling but not on a roof)\n3rd=[4] (includes speaking and some shoveling, but no roof)\n4th=[5] (roof work scene, but not snow-related)\n5th=[3] (clears snow from a car, irrelevant to roof context)\n</summary>\n</think>\n<answer> [2] > [1] > [4] > [5] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_Yg6jFRYIyXw.mp4", "activitynet/videos/v_ffyPN2-EDg4.mp4", "activitynet/videos/v_yslzj3NGuLU.mp4", "activitynet/videos/v_Zl6o69-R2DE.mp4", "activitynet/videos/v_NDWJOqVUGeA.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A girl is sitting on a large chair with two girls beside her and a man tattooing her foot. Another man steps into frame and the girls continue talking to each other while the artist.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A girl sits on a large chair with two girls beside her, and a man is tattooing her foot. Another man steps into the frame as the girls talk.\n[2] Close-up view of tattooing an arm, with no additional people or chairs visible.\n[3] A woman is outside talking, then getting a tattoo indoors on her arm, with no indication of multiple girls or a focus on feet.\n[4] A child receiving a tattoo, unrelated to the query scenario.\n[5] A girl is lying on a chair being tattooed on the foot by a single artist; doesn't match multiple girls interacting as much.\n\n</content>\n\n<contrast>\n- [1] matches the query by showing a girl on a chair, accompanied by two others, with a man tattooing her foot, while another man enters.\n- [2] lacks the context of multiple girls and chair setting, focusing solely on close-up tattooing.\n- [3] involves a woman but focuses on arm tattooing and lacks girls around a chair setting.\n- [4] involves a child and is irrelevant to the query.\n- [5] features a girl being tattooed on the foot, but without the full social interaction indicated in the query.\n\n</contrast>\n\n<summary>\n1st=[1] (matches scene with multiple girls, foot tattooing, social interaction)\n2nd=[5] (foot tattooing present, fewer contextual details)\n3rd=[2] (tattooing present, lacks contextual match)\n4th=[3] (visibility of tattooing, but arm focus and outdoor context)\n5th=[4] (irrelevant to query context, child tattooing)\n\n</summary>\n</think>\n<answer> [1] > [5] > [2] > [3] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_K-r_Tj7uFGw.mp4", "activitynet/videos/v_JzLqc6MhJtc.mp4", "activitynet/videos/v_AyYpsQ4xOo8.mp4", "activitynet/videos/v_ZazjI26IQUo.mp4", "activitynet/videos/v_-tv6laq1ic0.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman behind a bar is showing how to make a mixed drink. A series of shots appear with a person pouring alcohol into a martini glass and then the screen fades into the video title screen showing a website name, the name of the drink they are making and the person who will be making it. A website address is shown at the bottom of the screen. A girl standing behind the bar explains what ingredients she is using for the drink, the amounts she is using, and mixes them all together in a glass and presents the drink. The camera zooms in on the drink and fades to black to end the video.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video opens with a title screen, followed by a woman behind a bar explaining ingredients and making a drink. She pours alcohol into a glass, finishes making the drink, and the video ends with a camera zoom-in and fade to black.\n[2] This video begins with a title screen. A woman explains ingredients while preparing a drink, pouring various liquids into a glass. The drink is completed with a close-up, then fades out.\n[3] The video starts with the title screen, followed by a woman behind a bar. She explains the drink-making process, measures ingredients, and pours them into a glass. It ends with a close-up and fade to black.\n[4] This video follows the same pattern. It opens with a title screen, shows a woman behind the bar preparing a drink with explanations, and concludes with a close-up of the drink followed by a fade to black.\n[5] Starting with a title screen, a woman details the ingredients and preparation of a drink, pouring liquids into a glass and completing the drink. It ends with a close-up and fade to black.\n</content>\n\n<contrast>\n- [1] best matches the query because it shows a clear progression from introduction to explanation, mixing, and concluding with a close-up and fade out.\n- [4] closely follows, capturing all essential actions but is slightly less clear in presentation.\n- [3] follows well with a structured demonstration but lacks some clarity in close-up.\n- [2] shows a complete cycle but is slightly less clear on the instructional part.\n- [5] is similar but less cohesive, with less focus on drink presentation.\n</contrast>\n\n<summary>\n1st=[1] (best matches query: clear progression, close-up, fade out)\n2nd=[4] (strong elements but slightly less clear)\n3rd=[3] (structured but less close-up focus)\n4th=[2] (complete but less instructional clarity)\n5th=[5] (less cohesion and focus on presentation)\n</summary>\n</think>\n<answer> [1] > [4] > [3] > [2] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_TOfBNYOlv8c.mp4", "activitynet/videos/v_Rl6US0JizDs.mp4", "activitynet/videos/v_qOi_E_uYCIo.mp4", "activitynet/videos/v_83a5nOzZU1g.mp4", "activitynet/videos/v_sZf6ijFwPDs.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Two gymasts are performing in front of a crowd. One is spinning constantly on a beam. The man in red mounts and begins to spin, performing several tricks. The man dismounts, and puts one victorious arm up into the air.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A gymnast in red performs on parallel bars in front of a large crowd, executing various flips and spins, eventually dismounting and celebrating with an arm raised.\n[2] Displays a similar performance by a gymnast in red, involving spins and flips on the parallel bars, with the sequence completing in a landing pose.\n[3] Shows a gymnast in red also performing on parallel bars with spins and handstands, ending with the gymnast stepping down.\n[4] Features a performance by a gymnast wearing a red and black uniform, performing multiple spins and handstands on parallel bars, followed by a controlled dismount.\n[5] Focuses on a pommel horse performance in an Olympic setting with quick spins and flips, concluding with audience celebrations and medal presentations.\n</content>\n\n<contrast>\n- [1] beats [2] because it ends with a clear victorious pose, aligning closely with the query.\n- [2] beats [3] because of better visibility of the gymnast's full routine, including spinning.\n- [3] beats [4] because the routine is more complete and visible, whereas [4] has more obstructions.\n- [4] beats [5] because it involves parallel bars, matching the query's context better than the pommel horse routine in [5].\n- [5] is last due to a different focus on pommel horse and less relevant final action.\n</contrast>\n\n<summary>\n1st=[1] (Best matches the query with a victorious pose and clear routine completion)\n2nd=[2] (Good sequence visibility and matching routine elements)\n3rd=[3] (Complete routine, though less final action clarity)\n4th=[4] (Relevant bar routine but less clear sequence ending)\n5th=[5] (Focus on a different apparatus and medal ceremony)\n</summary>\n</think>\n<answer> [1] > [2] > [3] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_21krmbuT21s.mp4", "activitynet/videos/v_5kBKAfEX7XA.mp4", "activitynet/videos/v_XuSWH1lfcvI.mp4", "activitynet/videos/v_zLVNSgR4IoA.mp4", "activitynet/videos/v_Le6fNx5IWRI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A group of men are standing on a diving board and they take terms jumping off of the diving board. As the last person jumps off the board,a group of men are shown walking across the pool.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows multiple men standing on a diving board, taking turns jumping off into a pool. After the final jump, a group of individuals is seen walking across the pool area.\n\n[2] This video depicts a diving session at a pool where a few individuals take turns diving off a board. There are multiple jumps, but no clear sequence of a group walking across the pool area afterward.\n\n[3] A single person is seen repeatedly diving off the board in an indoor pool setting. There is no group of men taking turns or walking across the pool area.\n\n[4] High diving events show various individuals diving off different heights, but there is no sequence showing multiple people taking turns and then walking across the pool.\n\n[5] Multiple individuals are on a diving board taking turns to dive, later transitioning to a group of people walking across the pool area, which fits the query description well.\n</content>\n\n<contrast>\n- [5] matches the query as it shows men taking turns diving and walking across the pool, whereas [1] shows similar actions but lacks clarity.\n- [1] is ranked higher than [2] because it involves a sequence of men diving and walking, whereas [2] lacks the final walking scene.\n- [2] involves multiple dives but lacks the walking segment found in [1].\n- [4] involves diverse diving events but lacks a coherent sequence of group diving and walking.\n- [3] focuses on a single diver, lacking multiple participants and the walking scene.\n</contrast>\n\n<summary>\n1st=[5] (Complete match with men diving and walking.)\n2nd=[1] (Shows men diving and transition but less clear.)\n3rd=[2] (Shows multiple diving but lacks walking.)\n4th=[4] (Diverse diving events, lacks group sequence.)\n5th=[3] (Single person diving, no group sequence.)\n</summary>\n</think>\n<answer> [5] > [1] > [2] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_gGg-kio0dmU.mp4", "activitynet/videos/v_Tu1oKxyfHTs.mp4", "activitynet/videos/v_mnS-nMZne3I.mp4", "activitynet/videos/v_vcE8e-fQBhs.mp4", "activitynet/videos/v_3TbZ7uSYUe8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A close up of a globe is shown with a person pointing a hookah at it. The man is then seen looking closely to the camera. He finally blows smoke rings out of his mouth.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video starts with a close-up of a plasma globe with a person pointing a hookah hose at it. The person then looks closely at the camera and proceeds to blow a sequence of smoke rings out of their mouth.\n\n[2] This video shows a person interacting with a hookah in a seated environment, but there's no close-up of a globe. The person gestures and eventually blows smoke, but not specifically in smoke rings.\n\n[3] Features a person using a hookah and blowing smoke, including some rings, but lacks the sequence starting with a globe interaction.\n\n[4] Displays a person smoking a hookah with smoke exhaled, but there is no interaction with a globe or specific close-up actions mentioned.\n\n[5] The sequence involves smoking a hookah, but is set at night with environmental views and text about the surroundings, lacking the globe interaction and blowing of smoke rings.\n</content>\n\n<contrast>\n- [1] aligns closely with the query since it includes the globe interaction and visible smoke rings, whereas [2] and [3] lack the initial globe interaction.\n- [3] shows some smoke ring activity but lacks the globe interaction seen in [1], making it less fitting.\n- [4] involves smoking but does not match the specific query actions like globe interaction or visible smoke rings.\n- [5] focuses more on the setting and environmental aspects rather than the specific actions of the query.\n</contrast>\n\n<summary>\n1st=[1] because it perfectly matches both the globe interaction and smoke rings described.\n2nd=[2] because it shows hookah smoking and gestures, though missing the globe interaction and smoke rings.\n3rd=[3] because it includes some smoke rings but lacks the globe component.\n4th=[4] because it depicts regular smoking without the specific actions involving a globe or rings.\n5th=[5] due to focusing on environment and lacking key actions like the globe interaction and blowing smoke rings.\n</summary>\n</think>\n<answer> [1] > [2] > [3] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_xld0M1MXNiE.mp4", "activitynet/videos/v_jqRJM8b4pgE.mp4", "activitynet/videos/v_NLkJgnrKaKM.mp4", "activitynet/videos/v_2L9pp85KvbQ.mp4", "activitynet/videos/v_lCuMD23jOxo.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A young girl is seen standing on a tennis court holding a racket. The woman begins swinging the racket while another boy stands beside her. The woman continues swinging and walks back to the boy.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Features a young girl on a tennis court but she is not holding a tennis racket. The scene involves an indoor court with tennis balls scattered around, and a person beside her, but the action is more about observing tennis.\n[2] Shows a young girl in a pink top on a tennis court. She is swinging a racket, with a boy standing nearby. This closely matches the query about swinging a racket and walking back.\n[3] Displays a girl using a leaf blower on a driveway, which doesn't match the tennis theme.\n[4] Occurs in an indoor squash court with adults playing. This isn't relevant to the tennis-specific query.\n[5] Takes place in a gymnasium with children doing gymnastics or dancing activities, not involving tennis or a tennis racket.\n</content>\n\n<contrast>\n- [2] matches the query because it shows a young girl swinging a racket on a tennis court and walking back, whereas [1] is more about watching or being on a court without swinging.\n- [2] is relevant because it features the tennis court setting and action mentioned, whereas [3], [4], and [5] involve different sports or actions.\n</contrast>\n\n<summary>\n1st=[2] (Matches the action and setting described in the query)\n2nd=[1] (Related setting but lacks the swinging action)\n3rd=[4] (Involves racket sports but not relevant setting or action)\n4th=[5] (Not related to tennis or swinging a racket)\n5th=[3] (Unrelated to tennis or the action described)\n</summary>\n</think>\n<answer> [2] > [1] > [4] > [5] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_w6Avae5on_0.mp4", "activitynet/videos/v_hsJJbTqSQR4.mp4", "activitynet/videos/v_f-4PjtdFbP4.mp4", "activitynet/videos/v_vrMWYB0UuGM.mp4", "activitynet/videos/v_eQMT-WmxKM8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Several men are playing soccer in a gym. There are some spectators in the background. Another group of men are playing soccer. There are spectators in the background. Various scenes of soccer matches are shown with spectators.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Displays indoor soccer matches with multiple teams playing in a gymnasium setting. There are visible spectators in some frames, adhering to the query's requirements. The presence of different teams and actions being shown aligns well with the query.\n[2] Consists of broad views of indoor soccer matches in a gymnasium with visible spectators. The focus on matches and the presence of spectators matches the query well. However, the action details are less clear compared to [1].\n[3] Contains a mix of indoor soccer clips, but with a heavy overlay of text and graphics, obscuring the main content. The indoor soccer aspect is present but not as clear or coherent as other options.\n[4] Shows people playing soccer indoors. The setting is correct, but there is less evidence of spectators, which makes it a weaker match to the query.\n[5] Features indoor soccer with several teams, and spectators are visible in some frames. However, the quality and focus on action are not as strong as [1] and [2].\n</content>\n\n<contrast>\n- [1] beats [2] because it shows more detailed action close-ups with spectators clearly visible.\n- [2] beats [5] because it provides broader views with recognizable spectators, though with less action detail.\n- [5] beats [4] because it includes more clear instances of soccer and spectators.\n- [4] beats [3] because the soccer action is less obscured by overlaid graphics.\n</contrast>\n\n<summary>\n1st=[1] (Clear action shots with spectators; strong adherence to query)\n2nd=[2] (Visible sports action and spectators; broader views)\n3rd=[5] (Some presence of soccer action and spectators; moderate clarity)\n4th=[4] (Less evidence of spectators; action visible)\n5th=[3] (Obscured action due to graphics; weaker match)\n</summary>\n</think>\n<answer> [1] > [2] > [5] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_w-6Oo3bgms4.mp4", "activitynet/videos/v_bMDyOXygNPM.mp4", "activitynet/videos/v_cCISvaObBxI.mp4", "activitynet/videos/v_5qublDzEgsw.mp4", "activitynet/videos/v_wlYxVUJSJVI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Cake chemistry explaining how to do a at home experiment, They give all the ingredients needed and begin to explain what they will be doing. The start to give step by step instructions on how to mix the ingredients together. They give very detailed information about what to do before putting them in the oven. Once, they come out of the oven they explain the difference between all of the cakes and the different ingredients they did not put in each to see why the results were different.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A person demonstrates cake decoration but lacks detail on experimental procedures or ingredient variations. It is more focused on the frosting aspect rather than cake chemistry.\n\n[2] This video shows a step-by-step cake mixing and baking process but doesn't reveal clear ingredient comparisons or detailed experiment discussions. There are general mixing instructions without emphasis on experimental outcomes.\n\n[3] The video provides a recipe for gluten-free chocolate cake, offering a coherent baking process but doesn't focus on experimentation or comparative explanations of ingredient variations.\n\n[4] This video presents an experiment involving different cake ingredients, showcasing the resultant cakes with missing components. It explains chemical reactions and ingredient impacts, aligning with the query's experimental focus.\n\n[5] Displays a step-by-step guide on making a cake with specific ingredients used but lacks detailed comparative analysis of experimental results or exclusion of certain substances.\n</content>\n\n<contrast>\n- [4] beats [2] because [4] focuses on ingredient variations and showcases experiment results, whereas [2] lacks an experimental approach.\n- [3] is ranked lower than [4] because [3] doesn't explicitly show experiments or results, focusing on a single recipe.\n- [1] is below [3] due to missing element of baking instructions; it's more about decorating.\n- [5] is above [1] for showing ingredient preparation but lacks experimental detail, unlike [4].\n</contrast>\n\n<summary>\n1st=[4] (matches the query’s focus on experimenting with cake ingredients)\n2nd=[2] (provides step-by-step instructions without experimental focus)\n3rd=[5] (ingredient list and preparation, lacking detailed experiments)\n4th=[3] (recipe demonstration without comparative experiment context)\n5th=[1] (focuses on decoration rather than baking chemistry)\n</summary>\n</think>\n<answer> [4] > [2] > [5] > [3] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_oNc2lJ9hV9E.mp4", "activitynet/videos/v_1f_EQl4C700.mp4", "activitynet/videos/v_XCwPsQwzfiA.mp4", "activitynet/videos/v_MBouLt-hXDU.mp4", "activitynet/videos/v_2PCZkpF1_wU.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A person in blue pants is raking some fallen leaves from the ground. There is a black dog and a white dog jumping around the person raking the leaves. The white dog is trying to bite off the rake. Both the dogs are playing around the man raking the leaves, trying to bite the rake. The person is moving the rake in circles as the dogs are jumping around him.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The person is playing with dogs in a park, not raking leaves. The dogs are chasing a frisbee, not interacting with a rake.\n[2] The person is seen raking leaves, but there's no sign of any dogs in the frames.\n[3] A person is raking leaves with a brown dog nearby. The dog is playing but does not interact with the rake much.\n[4] This video shows a person vacuuming indoors with a small dog, no leaves or raking involved.\n[5] A person in blue pants is raking leaves while a black dog and a white dog interact, occasionally biting at the rake, matching the query description.\n</content>\n\n<contrast>\n- [5] is ranked higher because it perfectly matches the query with the presence of two dogs interacting with a rake while a person in blue pants rakes leaves.\n- [3] ranks lower as, while it involves raking and a dog, it lacks the presence of two dogs and the specific interaction with the rake.\n- [2] shows raking but no dogs are present, making it less relevant to the query involving dogs.\n- [1] involves playing with dogs but lacks any raking or interaction with the rake.\n- [4] is least relevant, showing an indoor scene with no raking or outdoor activity.\n</contrast>\n\n<summary>\n1st=[5] (perfect match: raking with two dogs interacting with the rake)\n2nd=[3] (raking with a dog, but doesn't fully match the interaction and number of dogs)\n3rd=[2] (raking visible but no dogs)\n4th=[1] (dogs playing, unrelated to raking)\n5th=[4] (indoor vacuuming scene, not relevant)\n</summary>\n</think>\n<answer> [5] > [3] > [2] > [1] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_a2Xwx37YbhE.mp4", "activitynet/videos/v_08wKlw4a_so.mp4", "activitynet/videos/v_aBr9bdWrl0I.mp4", "activitynet/videos/v_k5vE0ehf5TA.mp4", "activitynet/videos/v_0MtST08F2Xc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A person is seen running along the road wearing equipment and speaking to the camera. The man runs around using strings as well as speaking to the camera. The man continues speaking while more shots are shown of him running.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A person is demonstrating a walking belt, occasionally appearing outdoors but primarily speaking to the camera indoors. He uses belts and strings but does not run extensively.\n[2] Shows a person using spring-loaded stilts to run outdoors. The equipment is more of a novelty and includes outdoor running similar to the query but lacks speaking to the camera.\n[3] Features a person using jumping stilts in a parking lot area. The focus is on trying out the stilts rather than running or speaking to the camera.\n[4] This video includes a person running wearing a suit equipped with strings and also speaking directly to the camera. It matches the query by showing running with equipment and speaking to the camera.\n[5] Displays professional athletes during events, primarily focusing on track events without any direct speaking to the camera. It shows running but lacks the personal interaction.\n</content>\n\n<contrast>\n- [4] best matches the query by showing a person running with equipment while speaking to the camera, whereas [1] focuses more on a demonstration with minimal running.\n- [2] shows a different kind of running with equipment but lacks the aspect of speaking to the camera present in [4].\n- [1] contains some elements of speaking about equipment, but it isn't focused on running as much as [4].\n- [3], though involving equipment, is more about using jump stilts and lacks coherence with the query’s focus on running and speaking to the camera.\n- [5] focuses on professional track events without personal interaction or casual equipment demonstration as seen in [1] and [4].\n</contrast>\n\n<summary>\n1st=[4] (best coherence with query: running with equipment and speaking to the camera)\n2nd=[1] (some speaking about equipment but lacks running)\n3rd=[2] (running with equipment but not speaking to the camera)\n4th=[3] (focuses on jumping stilts, lacks speaking)\n5th=[5] (focuses on professional events, lacks personal speaking)\n</summary>\n</think>\n<answer> [4] > [1] > [2] > [3] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_fXgszraP2tc.mp4", "activitynet/videos/v_qZTAv1s_eBQ.mp4", "activitynet/videos/v_aKhD_bFB7MA.mp4", "activitynet/videos/v_ffWMjZCysO0.mp4", "activitynet/videos/v_hFtmkU7wdx4.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A pinata hangs from a tree, as a group of girls stand in front of it. One has a stick and is blindfolded. She hits the pinata, then walks away.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows a pinata hanging as a group of people stand around. One girl is blindfolded and guided with a stick, interacting closely with the pinata. However, she is not clearly shown hitting it and walking away.\n\n[2] Features a pinata hanging from a tree with a small child appearing with a stick, some frames showing interaction but lacks focus on a specific girl with a blindfold consistently.\n\n[3] Displays a pinata hanging from a tree, with one girl visibly blindfolded and holding a stick. She strikes the pinata and walks away, fitting the query well in terms of sequence and coherence.\n\n[4] A pinata event with children, but it lacks the depiction of one girl blindfolded, hitting and walking away. The focus is on various children, not matching the specific sequence sought.\n\n[5] Displays a pinata activity with children, one being blindfolded, but the actions and focus are not consistently on one girl hitting and leaving the scene with clarity.\n</content>\n\n<contrast>\n- Image [3] fits best as it clearly shows a girl being blindfolded, hitting the pinata, and walking away, matching the query sequence.\n- Image [1] involves similar elements but does not explicitly show the hitting and departing actions as clearly as [3].\n- Image [5] includes a relevant blindfolding action but lacks clear depiction of the sequence.\n- Image [2] presents pinata interactions but with less focus on a singular sequence fitting the query.\n- Image [4] is least relevant due to missing focus on the specified sequence and actions with a blindfolded girl.\n</contrast>\n\n<summary>\n1st=[3] (best sequence match: blindfolding, hitting pinata, walking away)\n2nd=[1] (involves a pinata and participants, but less clear on sequence)\n3rd=[5] (shows some related actions, but lacks clarity and sequence)\n4th=[2] (depicts a pinata event, less focus on desired sequence)\n5th=[4] (least relevant sequence match: pinata context but not the specific actions)\n</summary>\n</think>\n<answer> [3] > [1] > [5] > [2] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_WJ4tCIdAgVo.mp4", "activitynet/videos/v_ibHRKaD5MbM.mp4", "activitynet/videos/v_tjuYS-6UC_M.mp4", "activitynet/videos/v_RgeVltfPDXQ.mp4", "activitynet/videos/v_T98RJsOiQWc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"a man is seated, talking to a talk show host. They are animated as they speak to one another. They continue talking, then the host presents the man with a gift.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a talk show setting with two men seated and talking animatedly. They then move away from the seats towards a ping-pong table, engaging in a game. There's no presentation of a gift observed.\n\n[2] Features a man and a talk show host sitting and talking. The interaction is animated, and towards the end, the host presents the man with a gift on a small pillow, matching the query perfectly.\n\n[3] This sequence shows two people at a desk displaying and discussing shoes. There is no talk show setting or gift presentation involved.\n\n[4] Shows a man talking with a host, where they engage in arm wrestling. There is no gift presentation seen.\n\n[5] Displays three people sitting and talking in a talk show format. They are engaged and animated, but there is no visible gift presentation.\n</content>\n\n<contrast>\n- [2] fits the query perfectly with a talk show setting, animated conversation, and a gift presentation, whereas [1] diverges into a ping-pong game without a gift scene.\n- [2] shows the final scene of gift presentation which is absent in [1], [3], [4], and [5].\n- [3] involves product discussion (shoes) rather than a talk show format, whereas [2] maintains the setting and flow relevant to the query.\n- [4] transitions into an arm wrestling scene instead of a gift presentation, lacking the closing interaction described in the query, making [2] more suitable.\n- [5] contains an animated talk among hosts and guests, lacking the specific gift-giving moment present in [2].\n</contrast>\n\n<summary>\n1st=[2] (Perfect match: animated talk show, gift presentation)\n2nd=[1] (Talk show setting, animated conversation, but no gift)\n3rd=[5] (Talk show format, animated conversation, but no gift)\n4th=[4] (Talk show with interaction, but incorrect scene transition)\n5th=[3] (Product showcase not relevant to talk show/gift theme)\n</summary>\n</think>\n<answer> [2] > [1] > [5] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_0jvbBtMIA8k.mp4", "activitynet/videos/v_NnEUVHhtLZs.mp4", "activitynet/videos/v_sS1Zh8mqMOE.mp4", "activitynet/videos/v_86sxvTk3YEY.mp4", "activitynet/videos/v_RJpWgi0EaUE.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"An intro starts and begins showing a group of people. Various people perform a dance routine together with an instructor performing front and center. The routine continues with the people and instructor moving around with their arms and legs doing various movements to music.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows a group of people in a gym setting with an instructor leading a dance routine. The sequence shows coordinated movements, with the instructor at the front. This matches the query of a dance routine with an instructor performing front and center.\n\n[2] Displays a group performing a dance routine on stage with an instructor leading. The vibrant lighting and visible stage suggest a performance or class setting, which aligns with the query.\n\n[3] Features a close-up of an instructor with a group in a training environment, focusing on martial arts (capoeira) movements rather than a dance routine with an instructor front and center.\n\n[4] Shows people performing slow, synchronized movements, likely martial arts or Tai Chi exercises, rather than a dance routine with varied arm and leg movements.\n\n[5] Depicts a group of people following an instructor in a dance studio, engaging in varied dance movements with the instructor prominent and a routine involving coordinated steps, matching the query.\n</content>\n\n<contrast>\n- [1] ranks higher than [2] because it starts with an introduction and directly shows a dance routine in a gym with an instructor, while [2] appears more like a performance.\n- [1] is chosen over [5] because it maintains a clearer sequence focusing on the instructor in a prominent dance class setting.\n- [2] is preferred over [5] due to the dynamic stage setting indicating an organized routine, albeit less focused on the introduction aspect.\n- [5] is ranked over [4] as it clearly shows a dance routine led by an instructor whereas [4] indicates martial arts or exercises.\n- [4] is chosen over [3] since it at least involves group movement, while [3] is a more instructional and discussion-oriented segment on martial arts.\n</contrast>\n\n<summary>\n1st=[1] (Starts with intro, clear dance routine with instructor front and center)\n2nd=[2] (Visible dance routine on stage, less focused on intro)\n3rd=[5] (Shows a studio dance routine but lacks intro prominence)\n4th=[4] (Involves group movement but not dance)\n5th=[3] (Martial arts focus, not a dance routine)\n</summary>\n</think>\n<answer> [1] > [2] > [5] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_zDPsJ3ECaTI.mp4", "activitynet/videos/v_Z3bPr9odxpw.mp4", "activitynet/videos/v_h3uTJ7HknzA.mp4", "activitynet/videos/v_jLwUrFMRGzY.mp4", "activitynet/videos/v_8EGM3zcvjs8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A female news anchor introduces a story. A female news reporter speaks to a fitness instructor inside a gym, where cyclists are riding fitness bikes. Exercises such as yoga and bicycling are demonstrated by those in the gym. The female news reporter continues talking to the fitness instructor. She speaks to the anchor back in the studio.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Features a news segment: a female news anchor introduces the story, and a female reporter speaks to a fitness instructor inside a gym. The footage shows activities like yoga and cycling. Reporter interaction with the anchor concludes the scene.\n[2] Focuses on a fitness segment called \"Spin into Shape\" within a gym, showcasing various exercises on fitness bikes. No interaction with a studio anchor is visible.\n[3] Displays an advertisement for exercise equipment, featuring people using fitness machines indoors. It lacks any news anchor or reporter involvement.\n[4] Shows a video of a woman in a gym demonstrating exercises including floor workouts and cycling. There is no evidence of a news segment or anchor interaction.\n[5] Concentrates on cyclists in a gym setting, with close-ups of equipment and interviews, but lacks the interaction with a studio news anchor or a story introduction.\n</content>\n<contrast>\n- [1] beats [2] because it includes a visible news anchor and reporter interaction, matching the query's sequence.\n- [2] beats [5] because it features a structured fitness segment relevant to the query, though without anchor interaction, whereas [5] lacks news segment context.\n- [5] beats [4] because it has a more coherent gym setting relevant to cycling, though still lacking the anchor-reporter connection.\n- [4] beats [3] as it shows gym exercises, slightly aligning with the query theme, whereas [3] is a generic advertisement.\n- [3] is least relevant as it shows an advertisement, lacking context related to a news anchor or reporter.\n</contrast>\n<summary>\n1st=[1] (includes both a news anchor and a gym reporter, matching the query sequence)\n2nd=[2] (shows a structured fitness segment but lacks anchor interaction)\n3rd=[5] (shows gym cycling, but no bridge with news)\n4th=[4] (demonstrates exercises, less contextually relevant to the query)\n5th=[3] (focuses on exercise equipment advertisement, no relevance to news)\n</summary>\n</think>\n<answer> [1] > [2] > [5] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_QxsELAtsSWE.mp4", "activitynet/videos/v_nGsRDNIsc2o.mp4", "activitynet/videos/v_ClE77VhGQsk.mp4", "activitynet/videos/v_jwOZsKS25oc.mp4", "activitynet/videos/v_tgPjcWP7Vig.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A game of volleyball is being played between blue and yellow team. The yellow team servies the ball and it is hit continually back and forth until the yellow team scores a point. A player on the yellow team falls and the players on each team assist her. They take her shoe and sock off and massage her leg. Two people proceed to pick her up with the help of a third and take her off the court and lay her down to the side to continue care.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A volleyball game is played between blue and yellow teams. The yellow team serves, and actions are visible with the ball being hit back and forth. A player from the yellow team falls, and the players attend to her, removing her shoe and sock, massaging her leg, and then carrying her off the court.\n[2] Indoor volleyball scenes primarily show practice and interviews with coaches and players. There is no visible injury or specific interactions related to the described scenario.\n[3] A beach volleyball game is visible with players on sand courts. It focuses on gameplay and interactions but lacks any injury or caregiving scenario.\n[4] A beach volleyball event with multiple players and spectators. Gameplay and interaction scenes, but no specific injury or caregiving is shown.\n[5] A casual beach volleyball game in a city setting with no visible injuries or player assistance occurring.\n</content>\n\n<contrast>\n- [1] matches with the gameplay detail, injury, and assistance described in the query, whereas [2] is focused on indoor practice and lacks injury-related action.\n- [1] contains both the game action and player assistance, fulfilling the narrative of the query, whereas [3], [4], and [5] depict general volleyball games without the context of injury assistance.\n- [2] lacks a continuous game and injury narrative, unlike [1], which provides detailed sequences of both gameplay and injury assistance.\n</contrast>\n\n<summary>\n1st=[1] (Shows the game, injury, assistance as described)\n2nd=[3] (Contains extensive gameplay but no injury or assistance)\n3rd=[4] (Shows gameplay but lacks specific injury or assistance context)\n4th=[5] (Casual gameplay, no injury or assistance shown)\n5th=[2] (Focus on practice and interviews, no relevant gameplay or injury sequence)\n</summary>\n</think>\n<answer> [1] > [3] > [4] > [5] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_3gsF785TAmg.mp4", "activitynet/videos/v_RZqM8Z0ocjE.mp4", "activitynet/videos/v_tl2hDYGBfqU.mp4", "activitynet/videos/v_k67UWfSt-iE.mp4", "activitynet/videos/v_D8dHEAYIawg.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"We see an opening screen and a man runs and does a high jump at fifteen feet. We see a man jump the high jump on a track at various height with lots of cut scenes. We see seventeen and a still of the man. We see fifteen nine scroll across the screen. We see a series of still of the man and sixteen six. Seventeen two scrolls across the screen. We see seventeen nine scroll across the screen. We then see the ending credits.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a pole vaulter attempting jumps at various heights, marked by numbers such as \"Sixteen\" and \"Seventeen.\" It includes a series of action shots and the text \"Still going on to reach higher heights.\"\n\n[2] The video features a high jumper named \"Sotomayor\" with various attempts shown, including a focus on distances like \"2.45\" meters. Multiple scenes show the approach and jump.\n\n[3] Focuses on a track and field event highlighting attempts by athletes \"Craddock\" and \"Taylor\" with distances displayed in meters, mostly related to triple jump events.\n\n[4] Shows a high jumper, possibly named \"Ukhov,\" performing jumps with various heights displayed, alongside close-ups and audience reactions.\n\n[5] Displays historical long jump clips with athletes and their jumping distances, such as \"8.74 m\" by \"Larry Myricks,\" with dates and achievements noted.\n</content>\n\n<contrast>\n- [1] aligns closely because it shows pole vaulting with text indicating specific heights, similar to the query description.\n- [2] involves high jumping but with less focus on different heights and scrolling text.\n- [3] revolves around triple jumping rather than high or pole vaulting.\n- [4] depicts high jumping with different heights similar to [2] but lacks specific content details.\n- [5] centers on long jumping, focusing on distances rather than heights like pole vaulting.\n</contrast>\n\n<summary>\n1st=[1] (Direct pole vaulting sequence with heights marked)\n2nd=[2] (Relevant high jumping, less emphasis on different heights)\n3rd=[4] (High jumping shown but lacks query-specific scrolling text)\n4th=[3] (Triple jumping, unrelated to pole/high jump query)\n5th=[5] (Historical long jump focus, unrelated to pole/high jump heights)\n</summary>\n</think>\n<answer> [1] > [2] > [4] > [3] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_OH5UXwa3F7E.mp4", "activitynet/videos/v_VCg6_fuipp8.mp4", "activitynet/videos/v_tJLm2D4J3X0.mp4", "activitynet/videos/v_R3ffxWaI9QU.mp4", "activitynet/videos/v_VM-ldOw7e4Y.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A cheer-leading team begins to hold up posters as their mascot runs behind them. They then begin to do a routine and some of the girls run with streamers as the rest of them hold up the girls for their stunt. The stunt is finished and the retreat back to their original position and begin doing cartwheels and other tricks and stunts along with their mascot. As the routine continues,more stunts are performed and the Tiger mascot continues to cheer with the girls.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows cheerleaders performing various stunts indoors with no visible mascot and no posters being held up initially. They perform stunts such as throws and formations but lack the specific sequence described in the query.\n\n[2] This video features a cheerleading team performing indoors on a mat. The routine includes synchronized moves and stunts but lacks a mascot or any indication of poster usage.\n\n[3] The footage captures a cheerleading performance with multiple stunts, and a large group involved. The sequence is dynamic with various cheer formations but doesn't show a mascot interacting or posters.\n\n[4] This video showcases a cheer routine with visible stunts and a mascot in a tiger costume. The mascot is actively involved with the cheerleaders, aligning closely with the query's description.\n\n[5] The video shows outdoor cheerleading on a field with formations and flips. It lacks a mascot visible in the routine and doesn't indicate any posters, focusing on the cheerleading stunts instead.\n</content>\n\n<contrast>\n- [4] is most relevant because it includes a mascot interacting with the cheerleaders, which aligns with the query, whereas [3] has dynamic stunts but no mascot interaction.\n- [3] is more relevant than [1] due to its complexity and dynamic cheer formats, although neither shows a mascot interaction.\n- [1] has better detailed stunts than [5] despite lacking a mascot or poster elements.\n- [5] presents typical cheer activities but without mascot involvement, less relevant than [2] which has well-executed stunts.\n- [2] and [5] are similar in lacking a mascot, but [2] has better coverage of stunts with greater clarity.\n</contrast>\n\n<summary>\n1st=[4] (Includes mascot and cheer routine closely matching query)\n2nd=[3] (Complex cheer stunts but lacks mascot)\n3rd=[1] (Detailed stunts; lacks mascot and posters)\n4th=[2] (Clear stunt coverage without mascot)\n5th=[5] (Outdoor cheer routine lacking mascot)\n</summary>\n</think>\n<answer> [4] > [3] > [1] > [2] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_iaqFvTpoceI.mp4", "activitynet/videos/v_fZ_CmvF-_No.mp4", "activitynet/videos/v_IC61TQ-sd8o.mp4", "activitynet/videos/v_1fF1OQtpctg.mp4", "activitynet/videos/v_4-Xe_9Ywvd8.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Alexandra Raisman from Massachusetts is seen wearing purple leotards, getting ready to do gymnastics on the horse bar. She jumps right onto the horse bar and does some front and back flips. Then she does some back hand springs on the horse bar. she swiftly jumps on the horse bar while maintaining complete control over body and balance. She does a final back flip and a front flip and then, gets off the horse bar. The spectators cheer for her as she walks away.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The gymnast in purple leotards performs on the balance beam (mistakenly called horse bar). She does flips and shows controlled movements on the beam, and dismounts with a flip. The crowd is visible in the background. \n[2] The gymnast is wearing a different-colored leotard and performs on the balance beam. Includes flips and a final balancing pose.\n[3] The gymnast in a purple leotard performs various flips and moves on the balance beam, ending with a dismount and wave to the audience.\n[4] Similar performance in a purple leotard on the balance beam, with a focus on flips and balancing, ending with a dismount.\n[5] The gymnast wearing a purple leotard performs on the balance beam with multiple flips and movements, finishing with a dismount.\n</content>\n\n<contrast>\n- [1] is ranked higher because the gymnast wears purple leotards, matches the description better with detailed flips and a crowd.\n- [3] shows a gymnast in purple leotards with detailed flips but lacks crowd visibility as much as [1].\n- [5] shows a purple leotard performance on the beam but with less interaction and detail than [1].\n- [4] displays gymnastics in purple leotards but not as closely detailed or relevant as [1].\n- [2] has a different leotard color and less relevance to the query.\n</contrast>\n\n<summary>\n1st=[1] The gymnast wears purple, performs beam routines matching the query with visible crowd interaction.\n2nd=[3] Close match with purple leotard and flipping sequences, but less crowd interaction than [1].\n3rd=[5] Purple leotard performance with flips, but lower engagement or focus compared to [1].\n4th=[4] Flips and performance in purple leotards, but less detail in routines than [1].\n5th=[2] Different leotard color and less relevance than the others.\n</summary>\n</think>\n<answer> [1] > [3] > [5] > [4] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_62h9hnNTKWM.mp4", "activitynet/videos/v_xYM6h31PrM0.mp4", "activitynet/videos/v_56sQMnSUWaI.mp4", "activitynet/videos/v_bRVazNYfpGE.mp4", "activitynet/videos/v_sByOncVAMqg.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A news man is speaking in front of a group of people. A man is shown on a diving board high atop a cliff. He dives, jumping into the shallow waters below.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video captures various scenes related to an indoor swimming pool. It involves people diving into the pool and being judged, but it is not related to a newsman or a cliff dive.\n\n[2] This video predominantly shows a scuba diving experience. There are interactions on a boat and scenes in the water, but no newsman or cliff diving.\n\n[3] Similar to [2], this video involves underwater and scuba diving activities but also lacks any relation to a cliff dive and a newsman.\n\n[4] This video features divers in a competition setting with an audience, yet it lacks the key elements of a newsman and cliff diving into shallow waters.\n\n[5] The video contains a newsman speaking in front of a group, followed by a scene of a man on a diving board at the edge of a cliff. The sequence concludes with the diver leaping into the water below, matching the query closely.\n</content>\n\n<contrast>\n- [5] beats [1] because it specifically includes a newsman and a cliff dive, whereas [1] focuses on a pool setting with judges.\n- [5] beats [2] and [3] as they involve scuba diving activities, while [5] includes the cliff diving sequence essential to the query.\n- [5] beats [4] as it contains the newsman and cliff components, while [4] focuses on poolside diving events without these key elements.\n</contrast>\n\n<summary>\n1st=[5] (precisely matches a newsman speaking and a cliff dive)\n2nd=[1] (features diving and an audience, though indoors)\n3rd=[4] (focuses on competitive diving, lacks cliff aspect)\n4th=[2] (relates to diving, but underwater, not cliff-related)\n5th=[3] (similar to [2], focused on scuba, not cliff-related)\n</summary>\n</think>\n<answer> [5] > [1] > [4] > [2] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_-y5p8UMdM20.mp4", "activitynet/videos/v_YmAIoN4uuz0.mp4", "activitynet/videos/v_wfc_R5VnoHU.mp4", "activitynet/videos/v_L5nqeFWufrE.mp4", "activitynet/videos/v_4innF37ifyY.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"We see a GoPro opening screen. We see a lady walking with a raft into a river. We see people riding rafts in the river. We see a mans's legs and the lady rides past. We see the lady handing a beverage in a can to the cameraman. The cameraman goes under a branch that is across the water. We then see the GoPro ending scene.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video features scenes of people rafting, with visible helmets and rafts, and includes moments of interaction on the river. However, it doesn't show a specific opening or closing GoPro screen, or the detail of a beverage exchange.\n\n[2] Shows people tubing on a river with GoPro sticks visible. While related to river activities, there's no footage of a specific woman with a raft and the subsequent interactions mentioned in the query.\n\n[3] Contains scenes of intense white-water rafting with helmets and paddles but lacks the GoPro-specific segments, casual river floating, and social interactions described in the query.\n\n[4] Features river tubing and underwater scenes but does not include the GoPro framing, specific person, or precise sequence described in the query.\n\n[5] Begins with a GoPro opening screen, then follows a woman walking into a river with a raft. The footage includes scenes of rafting, the woman's interactions, a beverage exchange, and ends with a GoPro closing screen.\n</content>\n\n<contrast>\n- [5] beats [1] because it includes the exact opening and closing GoPro scenes and the specific actions described.\n- [5] is ranked higher than [2] because it accurately matches the GoPro opening, the woman with the raft, and detailed interactions.\n- [5] outperforms [3] by featuring the GoPro framing and specific social interactions.\n- [5] is above [4] due to containing the complete sequence involving the GoPro elements and the detailed actions.\n</contrast>\n\n<summary>\n1st=[5] (Detailed match with GoPro scenes and specified actions)\n2nd=[1] (Closest river activity match but lacks GoPro specifics)\n3rd=[2] (Similar river theme but missing key interactions)\n4th=[3] (Related rafting activities but lacks detailed sequence)\n5th=[4] (General river tubing without GoPro elements)\n</summary>\n</think>\n<answer> [5] > [1] > [2] > [3] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_EmlmKZ6Gi3Y.mp4", "activitynet/videos/v_8olnQLX0hO8.mp4", "activitynet/videos/v_r40TuTkt9y4.mp4", "activitynet/videos/v_hrN0hOmQrD4.mp4", "activitynet/videos/v_KpOSIph2hHw.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman holds the back section of a girls hair while she gets ready to style it. The woman uses a single pick comb to separate strands of hair. The woman clips back sections of hair. The woman pulls on strands of hairs to get them straightened out. The woman braids them the strands of hair together.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a woman styling a young woman's hair. The sequence includes holding hair, using a comb to separate strands, clipping sections, pulling strands to straighten, and braiding.\n\n[2] The video focuses on twisting hair into large braids, but lacks detailed visibility of small sections, comb separation, and initial styling steps.\n\n[3] The video involves a hairstyling session, but it focuses on a different style without showing combing or detailed braiding.\n\n[4] This video depicts cornrowing and securing extensions, involving actions such as parting and braiding, but not fully aligning with the query's specifics.\n\n[5] The video emphasizes brushing and straightening hair with a comb, but there's no evidence of sectioning, clipping, or braiding as described in the query.\n</content>\n\n<contrast>\n- [1] is more relevant because it clearly shows all required steps from the query, including combing, sectioning, and braiding, whereas [2] focuses on twisting without detailed sectioning.\n- [1] includes the initial and braiding phases, whereas [3] lacks comb separation and braiding actions.\n- [1] aligns with the query due to visible braiding and combing, whereas [4] focuses on cornrows with extensions.\n- [1] demonstrates comb usage and clear braiding steps, whereas [5] focuses solely on brushing and straightening.\n</contrast>\n\n<summary>\n1st=[1] (matches all steps: combing, sectioning, braiding)\n2nd=[2] (shows hair styling with some braiding)\n3rd=[4] (includes braiding but focuses on cornrows with extensions)\n4th=[3] (shows hairstyling but not detailed combing/braiding)\n5th=[5] (focuses on brushing, lacks braiding and sectioning)\n</summary>\n</think>\n<answer> [1] > [2] > [4] > [3] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_YZJghuPmff8.mp4", "activitynet/videos/v_uLT7PmD_AJA.mp4", "activitynet/videos/v_fM4pJeHb8hQ.mp4", "activitynet/videos/v_xbWAmySRE88.mp4", "activitynet/videos/v_H-hgVm5G54Y.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A news segment begins and black and white picture of a smiling girl appears on the screen and the banner on the screen says \"HENRICO TEEN DIES AFTER RACE FATHER RECOUNTS DAUGHTER'S FINAL MOMENTS\". The newscaster appears with her parents talking about the girl while clips of the race play, and still shot pictures of the girl appear on the camera. When the pictures and video clips of the interview and race are over, the newscaster is speaking alone back at the news station.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video features a news segment about a tragic incident where a Henrico teen dies after a race. It includes a newscaster, interviews with parents, race clips, and images of the girl, focusing on her final moments.\n\n[2] This video covers an elite polo event, showing the newscaster discussing the sport, various polo matches, and interviews.\n\n[3] The video features an interview discussing the historic inclusion of women in marathons, featuring archival footage and a speaker recounting events.\n\n[4] This video reports a field hockey player named Prince making a national team, featuring clips of games and commentary from the news anchor.\n\n[5] The video reports about shuffleboard, with historical footage and commentary from the newscaster, showing the sport’s history and current playing fields.\n</content>\n\n<contrast>\n- [1] is directly relevant to the query as it covers a news segment about a teen's death after a race, complete with interviews and relevant visuals of the race, fitting the query closely. \n- [2] focuses on elite polo, lacking relevance regarding a tragic race.\n- [3] discusses the inclusion of women in marathons but lacks the specific narrative of a recent tragedy.\n- [4] is about a field hockey player, unrelated to a race tragedy.\n- [5] covers shuffleboard, unrelated to the race event mentioned in the query.\n</contrast>\n\n<summary>\n1st=[1] (Directly matches the news segment theme and content mentioned in the query)\n2nd=[3] (Relevance to a running event, but not the specific race narrative)\n3rd=[2] (Sports event coverage, but unrelated to the query’s subject)\n4th=[4] (Sports-related, but field hockey is unrelated)\n5th=[5] (Different historical sport focus)\n</summary>\n</think>\n<answer> [1] > [3] > [2] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_hMUoeOFmquI.mp4", "activitynet/videos/v_QRcwvzq7QUM.mp4", "activitynet/videos/v_fOGXvBAmTsY.mp4", "activitynet/videos/v_MXN34b2OUjQ.mp4", "activitynet/videos/v_WBkqGBFsdgI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A close up of ballet shoes are shown followed by a woman putting them on and moving around in a bedroom. She moves her arms and feet in the point shoes and whips her hair all around the room.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video starts with a close-up of ballet shoes, followed by a woman in a bedroom putting on pointe shoes and continuing to dance. She moves her arms and feet, and there are moments where she twirls and whips her hair as described in the query.\n\n[2] The footage shows a young girl in a domestic setting doing various activities, such as picking items from the floor and moving around, but there's no connection to ballet or dancing in pointe shoes.\n\n[3] This video features a woman in a ballet studio demonstrating ballet moves at a barre. While similar in context to ballet, it lacks the specific setting of a bedroom and the described movements.\n\n[4] The focus is on a person putting on socks and tying shoelaces of casual shoes, with no relevance to ballet or dance.\n\n[5] The video involves someone cleaning and preparing sneakers, with no connection to ballet shoes or dancing activities.\n</content>\n\n<contrast>\n- [1] beats [3] because it shows a woman putting on ballet shoes and dancing in a bedroom, matching the query description closely, whereas [3] is set in a studio and lacks the same contextual details.\n- [3] beats [2] because it contains ballet-related content, whereas [2] involves unrelated activities.\n- [2] beats [4] because it shows more related movement and dynamics, whereas [4] is solely focused on putting on casual shoes.\n- [4] beats [5] because it involves footwear, albeit incorrect, whereas [5] completely lacks any ballet context.\n</contrast>\n\n<summary>\n1st=[1] (accurately matches the entire query sequence)\n2nd=[3] (related to ballet but in a studio setting)\n3rd=[2] (unrelated activities but shows movement)\n4th=[4] (shoe-related but not ballet)\n5th=[5] (no connection to ballet or dance)\n</summary>\n</think>\n<answer> [1] > [3] > [2] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_feio1hW6Dy0.mp4", "activitynet/videos/v_7ftCKS5SCCk.mp4", "activitynet/videos/v_ifDsS6ZRsRM.mp4", "activitynet/videos/v_hru57dsi0N0.mp4", "activitynet/videos/v_iJzt1rvGiOc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A ballerina is shown in class warming up her feet and walking away crying. Various actors talk to the camera about filming a movie about ballet and the moves they learned. The ballerinas are moving around the set while the camera is rolling. The actors continue talking about tricks and what they learned. A man plays the violin while the final scene is filmed and everyone is dancing.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a ballet class setting with a mix of interviews and practice scenes. A director and actors talk about ballet and film-making. There are scenes of ballet practice, discussions about moves, and possibly filming. The content matches well with the query details.\n\n[2] A series of clips features a girl band performing and dancing, both on stage and casually. The video lacks any relevance to ballet or filmmaking, instead focusing on pop music and dance performances.\n\n[3] This video features ballerinas performing and practicing on stage, capturing various moves and preparations. There is a strong emphasis on ballet but no clear evidence of filmmaking or actor interviews as suggested by the query.\n\n[4] Ballerinas are seen rehearsing in a studio. There is a focus on practice routines, but it lacks any indication of actor involvement or a filmmaking context, which the query specifies.\n\n[5] The video covers a casual day for individuals involved in sports or leisure activities like playing volleyball. It doesn't align with the ballet and filmmaking theme of the query.\n</content>\n\n<contrast>\n- [1] is most relevant because it includes scenes of ballet practice and interviews, directly matching the query's description of ballet, actors, and filmmaking, whereas [3] is more focused on ballet performance without the filmmaking aspect.\n- [3] ranks higher than [4] because it at least captures ballet performances, whereas [4] is limited to practice without filmmaker or actor involvement.\n- [4] is more related to ballet than [5], which does not focus on ballet, dancers, or filmmaking at all.\n- [5] lacks any related content to ballet or filmmaking compared to all other videos, which at least touch on these themes.\n</contrast>\n\n<summary>\n1st=[1] (Ballet and filmmaking context with interviews and practice)\n2nd=[3] (Focus on ballet performances but lacks filmmaking context)\n3rd=[4] (Ballet practice without filmmaking or actors)\n4th=[2] (Unrelated pop music and dance performances)\n5th=[5] (Casual sports and activities, no ballet or film context)\n</summary>\n</think>\n<answer> [1] > [3] > [4] > [2] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v__ekWWP0dQZM.mp4", "activitynet/videos/v_9XanCE5nX2Q.mp4", "activitynet/videos/v_li16LYa7us8.mp4", "activitynet/videos/v_-5Yp-vToI2E.mp4", "activitynet/videos/v_aAlbRFeu32E.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman is seen shaving down the fur of a dog while the camera shows many dogs sitting in crates. A woman is seen speaking to the camera and showing off several dogs with long hair and afterwards. The woman continues shaving down the sides of the dog while stopping to speak to the camera and showing off the shaved dog.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A woman is shaving a dog's fur, writing on a board, and showing grooming tools. There is focus on individual grooming rather than multiple dogs in crates.\n\n[2] Displays a grooming process in a facility setting with several dogs visible. There is a woman speaking to the camera, along with shots of dogs being shaved.\n\n[3] Features a dog grooming competition. Although relevant, it lacks focus on a single grooming action and the presence of dogs in crates.\n\n[4] Shows a woman grooming a single dog, primarily demonstrating the grooming process. No dogs in crates are visible.\n\n[5] Contains footage from a dog grooming event but primarily interviews and demonstrations without focus on a single grooming activity or dogs in crates.\n</content>\n\n<contrast>\n- [2] better matches the query with visible grooming activity and interaction, whereas [1] lacks multiple dogs and crates.\n- [3] lacks single grooming focus, whereas [4] shows individual grooming but no crates.\n- [5] does not show crate scenes or continuous grooming activity, similarly to [4].\n</contrast>\n\n<summary>\n1st=[2] (Includes grooming, speaking to camera, multiple dogs with some in cages)\n2nd=[1] (Shows grooming but lacks visible crates or multiple dogs)\n3rd=[4] (Grooming activity focused on one dog)\n4th=[3] (Grooming competition setting)\n5th=[5] (Event setting with interviews, lacks specific grooming focus)\n</summary>\n</think>\n<answer> [2] > [1] > [4] > [3] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_p8C7govzEPw.mp4", "activitynet/videos/v_dSsMVoPLQOk.mp4", "activitynet/videos/v_2fp9Ni7Ms6A.mp4", "activitynet/videos/v_u4nL6NXfsak.mp4", "activitynet/videos/v_63Zt34YuyRk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Ladies are standing in a workout room. The ladies are belly dancing. The ladies dip down and toss their hair backwards. They swing their dipped bodies. The ladies dip and bend backwards. The ladies throw their hair back and forth. 2:43 A man is in the room in the back and bends down. The ladies finish dancing and take a bow.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video displays multiple women performing a belly dance routine with flowing fabric. There are expressive movements, but no specific hair tossing or interaction with a male dancer is visible.\n[2] Contains women in costume performing various dances, including some hair movements. The presence of men is shown, but specific dance moves like dipping or bowing aren’t clear.\n[3] Women in a gym-like setting performing exercise routines. No distinct belly dancing, dipping, or male presence is evident.\n[4] Two women are dancing in a workout room and wearing belly dance costumes. Some movements involve dips and hair tosses, fitting the query. A man occasionally appears in the background.\n[5] A group dance routine in a studio setting. No belly dancing or clear hair-tossing is visible, nor is there a noticeable male dancer.\n</content>\n\n<contrast>\n- [4] fits best because it includes belly dancing, hair tosses, and a male figure in the background.\n- [2] is better than [1] because it features some relevant dance elements like hair movement and male presence.\n- [1] involves a similar dance form but lacks a male dancer or specific query actions.\n- [5] depicts a routine not aligned with belly dancing or query details.\n- [3] shows a workout setting unrelated to belly dancing.\n</contrast>\n\n<summary>\n1st=[4] (Contains belly dancing with relevant actions and a male presence.)\n2nd=[2] (Includes similar dance elements, partially matching the query.)\n3rd=[1] (Belly dancing is evident but lacks the male and specific actions.)\n4th=[5] (Lacks specific query details; no belly dancing.)\n5th=[3] (Workout setting unrelated to the query.)\n</summary>\n</think>\n<answer> [4] > [2] > [1] > [5] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_QQQ7jSVSG3A.mp4", "activitynet/videos/v_PxEu0AAH9LI.mp4", "activitynet/videos/v_KPv8ZGrb78Y.mp4", "activitynet/videos/v_HVfr9FnEj-Q.mp4", "activitynet/videos/v_nY8a1E4QH5Q.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Cotton balls are being set on fire and put out in various ways. An ax is used to put out the fire. A piece of wood is used to put out the fire.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video shows children hitting and chopping with axes against a piece of wood. No fire or cotton balls are present, making it irrelevant to the query.\n[2] Cotton balls and a piece of wood are set on fire and extinguished using another piece of wood, which closely matches the query.\n[3] The frames depict the creation of a campfire with kindling being ignited, unrelated to extinguishing fires with wood or an axe.\n[4] An individual uses an axe to chop wood into smaller pieces, but there is no evidence of fire or cotton balls being extinguished.\n[5] There is a person appearing to juggle or play with flames outdoors, but no cotton balls or extinguishing action is visible.\n</content>\n\n<contrast>\n- [2] matches because it directly involves cotton balls and wood being set on fire and potentially extinguished, whereas [1] lacks any fire-related activity.\n- [4] has axe use but lacks the fire extinguishing element present in [2].\n- [5] involves fire but not cotton balls or extinguishing as seen in [2].\n- [3] is unrelated as it shows starting a fire rather than extinguishing it.\n- [1] involves the use of wood but no fire, distinguishing it from [2].\n</contrast>\n\n<summary>\n1st=[2] (directly matches query: cotton balls and wood extinguishing fire)\n2nd=[4] (involves an axe, though no visible extinguishing)\n3rd=[5] (involves fire, but no extinguishing of cotton balls)\n4th=[3] (fire is started; no extinguishing action)\n5th=[1] (no relation to fire or extinguishing)\n</summary>\n</think>\n<answer> [2] > [4] > [5] > [3] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_wQbVxdVTN-I.mp4", "activitynet/videos/v_2kx--bbqXHg.mp4", "activitynet/videos/v_O337_rGYZPc.mp4", "activitynet/videos/v_Lg6QZST92xo.mp4", "activitynet/videos/v_3jb0w48ACkU.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman wash the head of a man and cuts his hair, then the woman stands next two man an touch the head of the young man. The woman cuts the hair of a young man using first a machine, and then a comb and scissors. Then, the woman dry the hair with a dryer and brush the head. After, the woman takes cream on her hands and comb the head of the man with her fingers to gives form to the hair. Next, the woman explains the hairstyle.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a woman washing and cutting a man's hair. She uses both clippers and scissors, followed by styling with a blower, brush, and cream. It ends with an explanation of the hairstyle.\n[2] This video displays a detailed sequence of cutting a man's hair using scissors and clippers, focusing on technique but not showing washing, styling with a dryer, or an explanation.\n[3] A young man has his hair cut using scissors and clippers, with a focus on cutting techniques. There is no washing, drying, or explanation segment.\n[4] The video involves cutting and styling a man's long hair, with emphasis on using fingers and a dryer for styling, but lacks specific hair washing and explaining the hairstyle.\n[5] A salon video featuring a stylist cutting and styling a woman's hair. The process includes washing and styling but is focused on women's hairstyles.\n\n</content>\n\n<contrast>\n- [1] provides a complete sequence aligning with the query, including washing, cutting, styling, and explaining whereas [2] lacks washing and explanation.\n- [3] focuses on cutting but misses washing and a final hairstyle explanation, unlike [1].\n- [4] involves styling techniques but does not include washing or explaining the hairstyle as comprehensively as [1].\n- [5] focuses on women's hairstyles and not on the detailed male haircut and styling process described in the query, unlike [1].\n\n</contrast>\n\n<summary>\n1st=[1] (Completeness: washing, cutting, styling, explanation)\n2nd=[2] (Clear cutting process, lacking washing and explanation)\n3rd=[3] (Focus on cutting techniques, lacking washing/drying/explanation)\n4th=[4] (Styling without washing or detailed explanation)\n5th=[5] (Not relevant: women's hair focus)\n\n</summary>\n</think>\n<answer> [1] > [2] > [3] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_lHy_OZ8GygM.mp4", "activitynet/videos/v_ZYAY3TEUjGI.mp4", "activitynet/videos/v_TT2e7-NbUxo.mp4", "activitynet/videos/v_hPQEkPZmNsU.mp4", "activitynet/videos/v_h9YbUJuYzLk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Two men plays foosball in an arena with two referees standing near by watching the fame. We see the bleacher an spectators. A point is scored and see people clapping and a replay and a title screen. The man in gray backs up and wipes sweat. We see the man in black speaking with the people behind him before the men return to the game. We see the ending title screen.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a table tennis match with two players in red and green shirts, diverse spectators, action replays, and logos. It misses the foosball-related content.\n\n[2] This video depicts two men playing foosball inside an arena setting. Referees are visible, spectators are present, a point is scored, and there’s interaction with the crowd. It fits the query perfectly, including elements like clapping and title screens.\n\n[3] Indoor soccer game in an arena with players in orange and black uniforms. Referees and spectators are seen, but no foosball content is present. \n\n[4] A table tennis match is shown with players in close competition. The audience is watching, but this does not match the foosball query.\n\n[5] The video captures two people playing foosball. However, it seems to be in a casual setting rather than an arena, lacking elements like referees or a crowd.\n</content>\n\n<contrast>\n- Video [2] fits the query best due to the presence of foosball, referees, spectators, and matching events.\n- [5] shows foosball but lacks an arena setting and additional contextual elements like referees or titles.\n- [1] and [4] involve table tennis, failing to address the foosball scenario.\n- [3] portrays soccer, lacking any foosball-related action. \n</contrast>\n\n<summary>\n1st=[2] (Perfect match with foosball game, referees, spectators, and events)\n2nd=[5] (Shows foosball but misses arena and formal elements)\n3rd=[1] (Table tennis match, irrelevant to foosball)\n4th=[4] (Another table tennis video with no relevance)\n5th=[3] (Indoor soccer game, not related to foosball)\n</summary>\n</think>\n<answer> [2] > [5] > [1] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_ahOnhEvMgq0.mp4", "activitynet/videos/v_0j9TxzwRr6E.mp4", "activitynet/videos/v_W34A9wUi-4g.mp4", "activitynet/videos/v_hJn5L1nrkL4.mp4", "activitynet/videos/v_cvatsuUhqhk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"There are two men playing table tennis in an indoor stadium. One of the players is wearing blue and the other player is in a red shirt. There is a person dressed in a black shirt seated near the players. There are several spectators seated in the stadium. They continue playing a rally for some time. The player in the red shirt misses the ball and gives his opponent a winning point. The player in blue cheers loudly after his winning point and breaks into a dance. He continues dancing and goes towards his opponent to mock him. He then goes back to the table tennis table. The man dressed on black shows the score.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video shows a foosball game instead of table tennis. Two men are engaged in playing foosball, and there are spectators, but this doesn't match the table tennis description.\n\n[2] This video features two men playing table tennis, but it's in a locker room setting rather than an indoor stadium. Players are not wearing blue and red, and there is no visible scoring or dancing.\n\n[3] The video shows a table tennis match in an indoor stadium, with players in red and another color, possibly green. A person in black is seated, and a rally is depicted. However, the sequence of events described in the query, such as dancing, is not visible.\n\n[4] There is a table tennis match in an indoor stadium setting. One player is in blue, and the other in red. A person in black is seated near the table, visible scoring, and one player celebrates a win by dancing. This closely matches the query details.\n\n[5] The footage includes people playing table tennis, but it's part of a show and doesn't have the indoor stadium or specific details regarding players' colors and actions.\n</content>\n\n<contrast>\n- [4] precisely matches the scene order and detail of the query, including the blue and red players, indoor setting, celebration, and scorer, whereas [3] misses the post-win actions like dancing.\n- [3] has elements like color and stadium setting but lacks sequence coherence compared to [4], where celebration and scoring match perfectly.\n- [5] features table tennis but lacks the indoor stadium setup and described actions found in [4].\n- [2] involves table tennis but is in a different setting without relevant actions.\n- [1] doesn't match as it’s not table tennis but foosball.\n</contrast>\n\n<summary>\n1st=[4] (matches the details of players, setting, actions, and scoring)\n2nd=[3] (matches indoor table tennis with similar player colors but lacks actions)\n3rd=[5] (involves table tennis but mismatched setting)\n4th=[2] (different setting and lacks match detail)\n5th=[1] (not table tennis)\n</summary>\n</think>\n<answer> [4] > [3] > [5] > [2] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_2Tm_2wgv2Zs.mp4", "activitynet/videos/v_6azNzt8rXAs.mp4", "activitynet/videos/v_ahOnhEvMgq0.mp4", "activitynet/videos/v_hJn5L1nrkL4.mp4", "activitynet/videos/v_G-zVpNXq224.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A close up of booklets are shown with a person flipping them open and reading. The woman marks a name down and peels it off to put on a sock. The woman then lays an iron over the name and peels off the paper to show how to properly label a gift.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Involves hand washing instructions for garments, showing a person washing fabric, no booklet flipping or labeling actions.\n[2] Focuses on nail art using paper clippings, no interactions with booklets, socks, or irons.\n[3] Shows a step-by-step process of labeling socks. A woman writes on a label, applies it to a sock, uses an iron, and removes the paper.\n[4] Demonstrates pumpkin carving with paper stencils, not related to booklets or labeling gifts.\n[5] Shows carpet installation process with no involvement of booklets or labeling.\n</content>\n\n<contrast>\n- [3] matches the query with visible actions: booklet flipping, labeling, ironing, and removing paper, whereas [1], [2], [4], and [5] lack these specific actions.\n- [1] focuses on garment washing, not labeling or iron usage like [3].\n- [2] involves nail design with no connection to labeling socks or gift preparation.\n- [4] revolves around pumpkin carving, lacking any labeling or iron use.\n- [5] is about carpet installation, unrelated to labeling or gift preparation actions described in the query.\n</contrast>\n\n<summary>\n1st=[3] (Perfect match: labeling socks process involving writing, ironing, and paper removal as described)\n2nd=[1] (Involves fabric-related tasks but no labeling or booklet-related actions)\n3rd=[4] (Includes stencil application but unrelated context of pumpkin carving)\n4th=[5] (Focus on carpet installation, no thematic relevance)\n5th=[2] (Concerned with nail art, unrelated to the query)\n</summary>\n</think>\n<answer> [3] > [1] > [4] > [5] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_QWFedtlPRYM.mp4", "activitynet/videos/v_0JCM9uzZafU.mp4", "activitynet/videos/v_JTGS1YulUQw.mp4", "activitynet/videos/v_a7QrIgqkyao.mp4", "activitynet/videos/v_mSPUMHD4F-E.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man is seen holding a glass while speaking to the camera when another man walks in and begins painting the wall. The man continues to paint up and down the wall while the camera captures him and the other man walks away.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows a man in a red shirt speaking to the camera, then applying wallpaper, but no man painting or walking away.\n[2] A man holds a glass while speaking to the camera, then another man paints a wall, fitting the query well.\n[3] Features a man speaking and wallpaper application but lacks a second man painting or leaving.\n[4] Displays a child and man interacting indoors, unrelated to painting or the described scene.\n[5] Includes a man speaking in a garden and another doing yard work, not matching the query about painting or indoor interaction.\n</content>\n\n<contrast>\n- [2] matches the query as it shows a man with a glass, another man painting, and coherent temporal action, whereas [1] and [3] involve wallpaper.\n- [2] fits better due to the described painting action, whereas [5] involves yard work unrelated to painting.\n- [3] and [1] involve speaking and decorating, but lack the second man and painting action as described.\n</contrast>\n\n<summary>\n1st=[2] (aligns with the query, shows a man speaking and another painting)\n2nd=[1] (man speaking, but involves wallpaper, not painting)\n3rd=[3] (man speaking and decorating, lacks painting)\n4th=[5] (features speaking and action, but in a garden)\n5th=[4] (unrelated indoor interaction)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [5] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_llcbN-s26vs.mp4", "activitynet/videos/v_GQVqSLs2Ulg.mp4", "activitynet/videos/v_Zso0ZBs2y6E.mp4", "activitynet/videos/v_F53s94562zc.mp4", "activitynet/videos/v_2Ua6E2a6kT4.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A lady pulls and fixes her outfit. The lady messes with her hair. The lady tries on shoes while standing and sitting.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] A person in a room is focused on trying on shoes, both standing and sitting, but there is no visible interaction with clothes or hair.\n[2] A lady is applying lipstick and touching her hair, but there is no scene of trying on shoes.\n[3] A person is adjusting shoes, with no visible interaction with clothes or hair.\n[4] A lady adjusts her outfit, manipulates her hair, and tries on shoes while standing and sitting, fulfilling all aspects of the query.\n[5] The focus is on a promotional setting for an ice cream brand, with no visible action related to clothes, hair, or trying on shoes.\n</content>\n\n<contrast>\n- [4] ranks higher than [1] because it shows a lady adjusting her outfit and trying on shoes, whereas [1] shows only shoe-trying actions.\n- [4] is better than [3] as it includes interactions with both outfit and hair, providing a full match to the query, whereas [3] shows only shoe adjustments.\n- [2] focuses on makeup and minor hair adjustments but lacks any shoe-trying actions, making it less relevant than [4].\n- [5] shows a commercial setup without any of the actions described in the query, making it the least relevant.\n</contrast>\n\n<summary>\n1st=[4] (Complete match of outfit adjustment, hair manipulation, and shoe trying)\n2nd=[1] (Involves shoe trying but lacks other actions)\n3rd=[3] (Involves shoe adjustment but lacks outfit and hair actions)\n4th=[2] (Shows some hair action but no shoe trying)\n5th=[5] (No relevant action for the query)\n</summary>\n</think>\n<answer> [4] > [1] > [3] > [2] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_9riZioI46sM.mp4", "activitynet/videos/v_avGe67wRvxA.mp4", "activitynet/videos/v_F2Igw6lvqwk.mp4", "activitynet/videos/v_RisNywzGXt8.mp4", "activitynet/videos/v_VNROQWtYhlw.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"An intro is shown of maps and a house close up. Afterwards a close up of a roof is shown as well as logos. More pictures are shown of trucks parked as well as logos.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows an intro with a logo and maps, followed by close-ups of a house and a roof, logos, and trucks parked, which matches the query well.\n[2] Focuses on an autoplex service, showing products, service area, and working on cars, which is unrelated to the query.\n[3] Displays logos and several roof close-ups, aligning well with the query, though no specific intro or parked trucks are shown.\n[4] Focuses on roof installation with various roof close-ups but lacks maps, house intro, or logos, partially aligning with the query.\n[5] Includes people on and working on roofs, close-ups of roofs, but lacks maps, logos, or parked trucks, partially aligning with the query.\n</content>\n\n<contrast>\n- [1] beats [3] because it includes maps, logos, and parked trucks, providing more comprehensive coverage of the query.\n- [3] beats [4] because it includes logos, aligning more closely with the query details.\n- [4] beats [5] because it provides more detailed coverage of roof close-ups and actions.\n- [5] beats [2] because it involves roofing, which is more relevant than autoplex services.\n</contrast>\n\n<summary>\n1st=[1] (matches all query elements: maps, house intro, roof, logos, and trucks)\n2nd=[3] (includes logos and roof close-ups but lacks comprehensive query details)\n3rd=[4] (detailed roof close-ups and actions but lacks maps and logos)\n4th=[5] (roof work present but lacks query's intro and parked trucks)\n5th=[2] (focuses on car services, unrelated to query)\n</summary>\n</think>\n<answer> [1] > [3] > [4] > [5] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_IT9uvR9yjxI.mp4", "activitynet/videos/v_MAUZbZnbM_Y.mp4", "activitynet/videos/v_0RZ5_8VD4vc.mp4", "activitynet/videos/v_ivBgqENFmt0.mp4", "activitynet/videos/v_-fMxoShIXiM.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"People run motocross on a bumpy road in the wood. A person climbs a bump with a motorcycle. Also people rides motorcycles on an unpaved flat surface.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows a motocross event with multiple riders on a dirt track, some on elevated terrains, but lacks a wooded environment. Includes promotional content unrelated to the query.\n\n[2] Features a young rider on a dirt bike across various surfaces including grassy and dirt areas; some bumps are visible. However, it primarily focuses on an individual rider and doesn't depict a wooded area or multiple riders on a defined motocross track.\n\n[3] Displays riders on a dirt and bumpy course, some riding up inclines. The setting includes natural surroundings that seem wooded. The videos show multiple motorcyclists and fits well with the query.\n\n[4] Contains motocross footage including stunts, aerial views, and various terrains but lacks the wooded environment specified in the query. Some videos show bikes on flat and bumpy surfaces but not in a forest.\n\n[5] Shows a motocross practice area with several motorcyclists on dusty and unpaved tracks, some of which are slightly elevated. Lacks wooded settings but shows multiple riders on a bumpy course, resembling part of the query situation.\n</content>\n\n<contrast>\n- [3] beats [1] because it includes wooded areas and focuses more closely on the bumpy road and climbing actions.\n- [3] beats [5] by providing a more diverse representation of wooded and bumpy road environments.\n- [5] ranks higher than [2] because it includes multiple riders on a defined track, which fits better with the query.\n- [1] beats [4] as it stays relevant to motocross events, while [4] includes unrelated stunts and promotional material.\n- [2] is less relevant compared to others as it focuses on one young rider and lacks a competitive motocross setting.\n</contrast>\n\n<summary>\n1st=[3] (wooded setting, multiple riders, bumpy terrain)\n2nd=[5] (captures bumpy surface and multiple riders, lacks trees)\n3rd=[1] (motocross event, lacks wooded area)\n4th=[4] (varied terrains but lacks setting relevance)\n5th=[2] (focuses on individual, lacks competitive setting)\n</summary>\n</think>\n<answer> [3] > [5] > [1] > [4] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_BfSqo9ONd6g.mp4", "activitynet/videos/v_UAgb3FEjBN8.mp4", "activitynet/videos/v_ux4iT6vQQb0.mp4", "activitynet/videos/v_Vq7Mm98I5Qs.mp4", "activitynet/videos/v_cA2ZLOZcCyc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A closeup of a bag of cement mix is shown. One man adds water into a container while another watches. The first man pours the cement mix into the container. The second man uses a mixer on the container. one of the men applies the mixture to a wall.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows men applying a mixture to walls but lacks key details like adding water, pouring ingredients, and mixing.\n\n[2] Focuses on tile laying and does not depict cement mixing or wall application.\n\n[3] Closely aligns with the query: starts with a closeup of a cement bag, then men add water, pour cement mix, use a mixer, and apply the mixture to a wall.\n\n[4] Displays wall rendering with cement but does not display critical steps like adding water or using a mixer.\n\n[5] Includes scenes of mixing and applying material to a wall, but lacks a closeup of a cement bag and clear depiction of adding water.\n</content>\n\n<contrast>\n- [3] fits best because it covers all key actions: closeup of the cement bag, adding water, pouring, mixing, and wall application, whereas [1] and [5] lack some of these elements.\n- [5] is closer to the query than [1] because it shows mixing and application, unlike [1] which mainly shows application.\n- [4] ranks lower than [5] as it misses several important steps like adding water and using a mixer.\n- [2] is least relevant, focusing on tile work with no relevant cement mixing or application actions.\n</contrast>\n\n<summary>\n1st=[3] captures entire process from bag to wall application.\n2nd=[5] includes mixing and application but misses cement bag closeup and adding water.\n3rd=[1] shows application but misses other steps.\n4th=[4] involves rendering but lacks setup steps like water and mixing.\n5th=[2] unrelated focus on tile work.\n</summary>\n</think>\n<answer> [3] > [5] > [1] > [4] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_lZ2X-e33E0A.mp4", "activitynet/videos/v_oP77DgsbhKQ.mp4", "activitynet/videos/v_JBTiD2-FcR0.mp4", "activitynet/videos/v_IWXIEfaltJ4.mp4", "activitynet/videos/v_SmBEf-g82Ew.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman is seen speaking to the camera and leads into the news interviewing a man. Several shots are then shown of two bulls fighting other while others watch and speak to the camera.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The sequence starts with a woman speaking to the camera, likely a news presenter. This is followed by an interview with a man. The later frames depict bulls engaging in combat with onlookers present, and another interview is conducted.\n\n[2] Features several scenes of bulls in an open area with crowds around, but lacks any interview or presenter seen speaking to the camera.\n\n[3] Consists of visual and textual content advocating against bullfighting, with no interviews or scenes of bullfighting occurring live, only static images and text.\n\n[4] Shows scenes from a rodeo event with interviews, but the focus is on roping events rather than bullfighting.\n\n[5] Contains images and text criticizing bullfighting, with no interviews or live combat scenes.\n</content>\n\n<contrast>\n- [1] is selected because it fully depicts a sequence where a woman speaks to the camera, followed by bullfighting and interviews, matching the query requirements.\n- [2] shows bullfighting but misses the introductory section with a woman speaking and interviews.\n- [3] and [5] lack a temporal sequence with interviews or live bullfighting captures.\n- [4] is irrelevant as it showcases roping, not bullfighting.\n</contrast>\n\n<summary>\n1st=[1] (Sequence matches query: woman speaking, interviews, bullfighting scenes)\n2nd=[2] (Shows bullfighting but lacks intro/interviews)\n3rd=[4] (Features interviews but focused on roping)\n4th=[3] (Text/images against bullfighting, lacks interview)\n5th=[5] (Text/images against bullfighting, lacks interview)\n</summary>\n</think>\n<answer> [1] > [2] > [4] > [3] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_i9g2MGNUVQI.mp4", "activitynet/videos/v_mb2fWzNSob0.mp4", "activitynet/videos/v_AdFg_21LRg4.mp4", "activitynet/videos/v_cufmgoo7E_k.mp4", "activitynet/videos/v_vpvhtd7CqUA.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman is shown talking to the camera with her hands and the camera pointing to a building. She stands in front of the building doing a several little dances and continues this in front of several buildings and locations.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Displays a woman outdoors at a park, but she isn't speaking to the camera or near buildings. Instead, she walks and poses near some hopscotch markings.\n\n[2] Shows various scenes with people in urban environments, some rollerblading and some casually talking to the camera. It features multiple individuals, not focused on one woman.\n\n[3] Shows a woman in a gym setting. She talks to the camera and uses an exercise bike, but there's no interaction with buildings or dancing.\n\n[4] Features a woman inside, talking to the camera briefly, but mostly interacting with others in a social setting. No outdoor building scenes or dancing.\n\n[5] Displays a woman walking and dancing in front of various coffee shops and buildings. She gestures while talking to the camera, fitting the query description well.\n</content>\n\n<contrast>\n- [5] provides a clear sequence where the woman talks to the camera and dances in front of buildings, closely aligning with the query, whereas [1] has no building interaction or direct camera address.\n- [3] involves indoor gym activities with no changing locations or buildings, unlike [5]'s relevance to different outdoor settings.\n- [2] includes various people and scenes, lacking the focused narrative of a single woman engaging with multiple buildings, as seen in [5].\n- [4] has brief camera interaction indoors, without the outdoor, building-focused scenario of [5].\n</contrast>\n\n<summary>\n1st=[5] (woman dances and talks in front of buildings)\n2nd=[4] (some camera interaction, but mostly indoors)\n3rd=[2] (urban scenes, lacks focused narrative)\n4th=[1] (outdoor activity, no camera speaking or building focus)\n5th=[3] (indoor gym, unrelated to query's outdoor setting)\n</summary>\n</think>\n<answer> [5] > [4] > [2] > [1] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_pieT6K771WQ.mp4", "activitynet/videos/v_JRr3BruqS2Y.mp4", "activitynet/videos/v_bGDwE1v16HU.mp4", "activitynet/videos/v_XFySa1Gc-yM.mp4", "activitynet/videos/v_igwT-3gprOI.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A person is seen standing on a court holding a tennis racket. The man then demonstrates several moves with the tennis racket and hitting the ball to another. The two continuously hit the ball back and fourth to one another.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows two people on a squash court holding rackets. They demonstrate movements but are not on a tennis court, and the sport appears to be squash, not tennis.\n\n[2] This sequence takes place on a tennis court. It includes a person holding a tennis racket and demonstrating various tennis moves, such as hitting the ball, aligning closely with the query.\n\n[3] The video depicts a person on a tennis court demonstrating tennis serves and movements, matching the description of demonstrating moves and hitting a tennis ball back and forth.\n\n[4] This sequence shows two people on a tennis court. They are holding tennis rackets and appear to be demonstrating techniques, fitting the query's requirements of a demonstration on a tennis court.\n\n[5] The activity occurs on a badminton court with players holding badminton rackets, which does not fit the query for tennis actions.\n</content>\n\n<contrast>\n- [2] ranks higher than [1] because it is on a tennis court and matches the query, whereas [1] involves squash.\n- [3] beats [4] due to clearer action demonstration on a tennis court, whereas [4] involves more static instruction.\n- [2] is superior to [3] as it shows more dynamic play and interaction fitting the query better.\n- [4] is higher than [5] due to being on a tennis court, whereas [5] involves badminton.\n- [5] is the least relevant as it involves a different sport.\n</contrast>\n\n<summary>\n1st=[2] (complete tennis demonstration matching the query)\n2nd=[3] (clear tennis actions and demonstration)\n3rd=[4] (instruction and demonstration on a tennis court)\n4th=[1] (squash rather than tennis)\n5th=[5] (features badminton, not tennis)\n</summary>\n</think>\n<answer> [2] > [3] > [4] > [1] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_xrbKvttgcMA.mp4", "activitynet/videos/v_0JpULUYPDX0.mp4", "activitynet/videos/v_BryW4niZT9Y.mp4", "activitynet/videos/v_nt4Ag91306U.mp4", "activitynet/videos/v_2Peh_gdQCjg.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"People are playing a game in a field. The man with the mallet and the man with the red hat run back and forth. A man with a blue shirt runs toward home plate. The man with the mallet is shown getting hit in slow motion. A woman is up to bat and is shown running after getting a hit.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] People are in a yard doing various activities like hitting with a mallet and swinging. There's no clear game or running as described in the query.\n[2] People are playing croquet, using mallets on a lawn. There's no evidence of running or a woman batting.\n[3] The video shows a field with people playing a game, some with paddles. There is running and potential batting, but not clear from the images if it matches the query details exactly.\n[4] Clips of a cricket game are shown, but this does not fit the query of running and batting in the manner described.\n[5] People are in a field playing a game with visible running, a man in a red hat, and a woman batting, closely matching the query sequence and participants.\n</content>\n\n<contrast>\n- [5] beats [3] because it shows a man with a red hat, a woman batting, and running, matching the query sequence specifically, whereas [3] has unclear relations to the query.\n- [3] beats [4] because it involves people on a field, potentially with paddles, whereas [4] shows cricket, which is unrelated.\n- [4] beats [2] because the player's actions are more dynamic, though not matching the query, whereas [2] is entirely focused on croquet actions.\n- [2] beats [1] because at least it involves a game context, whereas [1] is a mix of unrelated activities.\n</contrast>\n\n<summary>\n1st=[5] (Matches all aspects: field game, running, man with red hat, slow-motion, and a woman batting)\n2nd=[3] (Field game and some relevant motion)\n3rd=[4] (Cricket game dynamics, though not matching query)\n4th=[2] (Croquet, not related to the query)\n5th=[1] (Unrelated activities, not matching query actions)\n</summary>\n</think>\n<answer> [5] > [3] > [4] > [2] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_UYUvjnlU9Ho.mp4", "activitynet/videos/v_TpURw9nc8ks.mp4", "activitynet/videos/v_Lt6waJFJWI0.mp4", "activitynet/videos/v_U0IJmOHoa1M.mp4", "activitynet/videos/v_zlOErVVHKEk.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman is seen kneeling into frame next to a patio set. She then shows off various cleaning tools an rubs them along a chair. She brings the chair inside and pours paint into a bucket and paints along the side. She shows off the chair outside again while speaking to the camera.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The video shows a woman demonstrating how to refurbish a plastic patio chair. She starts by cleaning and prepping the chair with various tools, paints it, and finally displays the finished result outside while speaking to the camera. This closely matches the query description.\n\n[2] The video features a woman painting a small wooden table indoors. She discusses tools and paints while demonstrating the process. While related to painting furniture, it does not match the specific sequence and setup of the query.\n\n[3] The woman is engaged in refurbishing a wooden fence, not a patio chair. She shows tools and applies paint, but the action focuses entirely on outdoor fence work, diverging from the query's chair-related tasks.\n\n[4] A woman is seen refurbishing wooden cabinet doors, sanding, and painting. Similar painting activity is involved, but it pertains to cabinetry and lacks the scenario described in the query.\n\n[5] Displays the process of painting a rattan chair with close-up details of the work. However, it does not match the plastic chair or related actions in the query.\n</content>\n\n<contrast>\n- [1] is more relevant because it closely follows the query's sequence and actions, focusing on a plastic chair.\n- [2] while related to furniture painting, lacks the specific elements of chair and outdoor setup.\n- [3] focuses on fence painting, irrelevant to the chair-focused query.\n- [4] involves cabinetry work, diverging from the patio chair narrative.\n- [5] covers a different type of chair and lacks the query's details on the specific actions and outdoor setting.\n</contrast>\n\n<summary>\n1st=[1] (matches query sequence and elements)\n2nd=[2] (involves furniture painting but different setup)\n3rd=[5] (involves chair painting but different material)\n4th=[4] (cabinet painting, not a chair)\n5th=[3] (fence painting, irrelevant to query)\n</summary>\n</think>\n<answer> [1] > [2] > [5] > [4] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_xpEbejnuGEY.mp4", "activitynet/videos/v_cKHdMUfKsQ8.mp4", "activitynet/videos/v_LSvjzZHbetU.mp4", "activitynet/videos/v__JE5T2RKZvo.mp4", "activitynet/videos/v_I5YCAQXHoFw.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"There are lots of adults in the room (elderly or young) and they're all watching this girl play the clarinet. First the girl walks to the front of the classroom as she holds her clarinet and the man who plays the piano gets up from his seat to play the piano. Then he starts playing the piano and she plays the clarinet along with the piano, when she's done everyone claps for her.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The sequence features a man and a woman with a flute, not a clarinet. It doesn't match the query's narrative about a girl with a clarinet and a man playing the piano in front of an audience. \n\n[2] Shows a girl playing the flute in front of people who appear to be listening, some of whom are adults. This matches part of the performance aspect, with an audience and instrumental performance, though it's a flute, not a clarinet.\n\n[3] Displays a man with a saxophone demonstrating something in front of a piano. There are no scenes of audience or a girl playing, hence not matching the query.\n\n[4] A girl plays an accordion in a room with adults. There is an audience but not a clarinet or a piano. The narrative elements do not match the query.\n\n[5] The video is in black and white, showing a performance with a piano but no specific focus on a girl or a clarinet. There's an ensemble performance with an audience, but it lacks the specific elements described.\n</content>\n\n<contrast>\n- [2] fits more closely because it shows a young girl playing in front of an audience with a piano present, whereas [1] has a different instrument and interaction.\n- [2] features a girl and an audience, giving partial match, while [3] entirely lacks the presence of a girl or relevant instruments.\n- [4] has an audience but lacks the clarinet and piano elements, while [2] has a performance setting aligning with some aspects of the query.\n- [5] involves a group setting without focus on an individual girl with a clarinet, whereas [2] does present an individual performance setup.\n</contrast>\n\n<summary>\n1st=[2] (matches elements of performance with an audience and young girl, close in context)\n2nd=[4] (has an audience and music, though different instruments)\n3rd=[5] (performance setting but lacks individual focus as per query)\n4th=[1] (interaction with incorrect instruments for query)\n5th=[3] (lacks the girl and audience involvement)\n</summary>\n</think>\n<answer> [2] > [4] > [5] > [1] > [3] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_AYhduTuI_zM.mp4", "activitynet/videos/v_vpu4Bwq-xHI.mp4", "activitynet/videos/v_vWEbHvnkogA.mp4", "activitynet/videos/v_cErtrZrWhiQ.mp4", "activitynet/videos/v_eOmL8f-yB40.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A group of young Asian men are indoors with a lot of spectators in the stands holding signs and cheering them on. Two young Asian men stand up with their bow and arrows and they each take turns at shooting at their bulls eye target while the crowd cheers them on after each shot is made, and occasionally 3 suited people are shown making commentaries. When the men are done shooting at the target,one of them raise their arm in victory, they shake each others hands,hug, take a bow and wave to the cheering crowd.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The sequence shows a darts event with close-ups of the players, the dartboards, and the crowd. The setting is indoors with spectators cheering. No archery or Asian participants are visible.\n\n[2] This video depicts a darts match with close-ups of the players, dartboards, and audience reactions. Players are in dart attire, and the event is indoors with cheering, but no archery or Asian participants.\n\n[3] The video captures an archery competition indoors. Young Asian men shoot arrows at targets, with spectators cheering. The sequence includes archery action, target shots, and audience reactions, aligning well with the query.\n\n[4] This shows a high school archery team practicing indoors. It features young archers with bows and arrows aiming at targets, but lacks the detailed interaction and spectator engagement described in the query.\n\n[5] The footage focuses on a traditional outdoor archery demonstration with older participants. There is no large indoor audience or specific match ending celebration as described in the query.\n</content>\n\n<contrast>\n- [3] matches the query with young Asian participants in an indoor archery context and clear spectator reactions, whereas [1] and [2] feature darts.\n- [4] involves archery but lacks clear spectator involvement and specific victory actions, whereas [3] involves cheering and victory gestures.\n- [5] shows traditional archery without the competitive, celebratory environment described, whereas [3] depicts a lively competitive archery event.\n</contrast>\n\n<summary>\n1st=[3] (Matches query with indoor archery, young Asian men, spectator engagement)\n2nd=[4] (Features archery, lacks detailed ending/spectator interaction)\n3rd=[5] (Displays archery, lacks indoor/spectator context)\n4th=[1] (Darts event, incorrect sport)\n5th=[2] (Darts event, incorrect sport)\n</summary>\n</think>\n<answer> [3] > [4] > [5] > [1] > [2] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_nnWJGghixr0.mp4", "activitynet/videos/v_cjS2B52jBbU.mp4", "activitynet/videos/v_CDWHstJThJA.mp4", "activitynet/videos/v_dh4bxvmrCNs.mp4", "activitynet/videos/v_7lwrP2ma2TU.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"Two men do maneuvers on a canoe pivoting the canoe helping with the paddlers. Then, the men pivot the canoe away from the bow paddlers while paddling fast the oars. After, the men slides the canoe forward using the oars to advance. Next, the men sail on the river until arrive to the bank of the river.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The sequence shows two people on a canoe performing tandem maneuvers, such as pivoting and sliding the canoe sideways. It appears instructional but lacks significant action involving rapid paddling or arriving at a riverbank.\n\n[2] Displays two men paddling down a narrow river in a canoe, suggesting movement, although it lacks specific maneuvering such as pivoting or sliding.\n\n[3] Involves various water activities like canoeing and snorkeling. There is movement along a river, but no clear depiction of specific canoe maneuvers or arrival at a bank.\n\n[4] Shows a person navigating a canoe in rapids, demonstrating paddling skills but does not show tandem maneuvers or destination arrival.\n\n[5] Focuses on a single person in a canoe paddling across open water. No tandem maneuvers or rapid movements are depicted.\n\n</content>\n\n<contrast>\n- Video [1] shows tandem maneuvers pivoting the canoe, making it closely match the query, whereas [2] shows movement without specific maneuvers.\n- [2] beats [3] because it has two men in a canoe, aligning better with the query, while [3] includes diverse activities.\n- [3] over [4] as it involves a river, while [4] is primarily on rapids, lacking tandem interactions.\n- [4] is preferred over [5] because it involves dynamic paddling, fitting the theme better than solo paddling in open water in [5].\n- [5] ranks lowest due to focusing on solo paddling without maneuvering or riverbank arrival.\n\n</contrast>\n\n<summary>\n1st=[1] (Shows tandem canoe maneuvers closely aligned with the query)\n2nd=[2] (Shows two men paddling down a river, consistent with the theme)\n3rd=[3] (Includes river travel but lacks clear tandem maneuvers)\n4th=[4] (Depicts paddling skills but in rapids, not tandem)\n5th=[5] (Focuses on solo paddling without relevant tandem activities)\n\n</summary>\n</think>\n<answer> [1] > [2] > [3] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_NV2scxevYKk.mp4", "activitynet/videos/v_ABCRXwB5-JQ.mp4", "activitynet/videos/v_djgM0IWIBKc.mp4", "activitynet/videos/v_-IkL9Kq9zfQ.mp4", "activitynet/videos/v_BD_ON66t6rM.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A bag of Spill Master Absorbant is shown on the screen. A man stands in a warehouse talking to the camera and pours some liquid on the ground. He takes a bowl of some material and sprinkled it over the spills. He then sweeps up the material with a broom and dust pan.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Shows a construction scene with a worker on a rooftop, not related to the query involving Spill Master Absorbent or cleaning a spill.\n\n[2] Demonstrates vacuum cleaning with a focus on attachments and vacuuming techniques, not related to the query about using absorbents in a warehouse.\n\n[3] Focuses on stone and adhesive application for décor, unrelated to the spill cleanup context described in the query.\n\n[4] A man in a warehouse pours liquid on the ground, applies an absorbent material, and then sweeps it up, matching the query about using Spill Master Absorbent.\n\n[5] Features a shoe store with a person handling and discussing shoes, which is unrelated to the liquid spill and cleanup process described in the query.\n</content>\n\n<contrast>\n- [4] clearly matches the query as it involves pouring liquid, applying an absorbent, and demonstrating cleanup in a warehouse, whereas others do not depict this.\n- [2] involves cleaning but focuses on vacuuming rather than the use of an absorbent in a warehouse.\n- [3] shows construction work with stone and adhesive application, not related to spill cleanup.\n- [1] is about rooftop work unrelated to cleanup processes.\n- [5] focuses on shoes and product presentation, completely unrelated to spill absorption and cleanup.\n</contrast>\n\n<summary>\n1st=[4] (matches the entire sequence of actions described in the query)\n2nd=[2] (involves cleaning, though not with absorbents)\n3rd=[3] (shows material application, different domain)\n4th=[1] (construction setting, unrelated actions)\n5th=[5] (product presentation in a shoe store, entirely unrelated)\n</summary>\n</think>\n<answer> [4] > [2] > [3] > [1] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_NDWJOqVUGeA.mp4", "activitynet/videos/v_K8ZPkLVlAiA.mp4", "activitynet/videos/v_wr_1p7obP-g.mp4", "activitynet/videos/v_UxlDSfC7S-0.mp4", "activitynet/videos/v_KuyaO3brwBc.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A woman is video taping some men in a living room getting jackets on. One of the men motions the the camera person to follow him as he grabs a broom and walks down a hall and out a door to a snow covered parking lot. They walk out into the parking lot as another man comes outside. The man with the broom brushes snow off of an SUV as another man does the same at a car next to him.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] This video shows people indoors with no clear depiction of a living room scene of dressing. There are clips of night snow activities, but without specific actions like brushing off an SUV.\n[2] This video focuses on a man outside shoveling or demonstrating snow removal from cars. No indication of living room or following anyone.\n[3] This video starts indoors in a living room-like setting, showing people putting on jackets. A man directs the camera, grabs a broom, proceeds outside to a snowy parking lot, and begins clearing snow from vehicles, matching the full sequence.\n[4] This video shows a person outdoors brushing snow off cars and using a shovel. The indoor scene or initial actions are absent.\n[5] The video involves a woman discussing snow-related topics with some outdoor clearing activity shown, lacking the full sequence of events in the query.\n</content>\n\n<contrast>\n- [3] is best because it shows an indoor living room scene followed by a transition to outdoor snow removal, closely matching the query, whereas [1] lacks the entire sequence.\n- [3] beats [4] because it includes both indoor and outdoor actions mentioned, while [4] only shows outdoor snow removal.\n- [3] is superior to [2] due to having an initial indoor transition before outdoor actions; [2] is entirely outdoors.\n- [3] surpasses [5] because [5] lacks the full sequence and actions specific to brushing snow off multiple cars.\n- [2] ranks higher than [4] because it focuses more on snow removal specifically from cars, albeit without the indoor start.\n</contrast>\n\n<summary>\n1st=[3] Has the complete sequence of events from inside a living room to a parking lot snow activity.\n2nd=[2] Features outdoor snow removal from vehicles, lacking indoor preparation.\n3rd=[4] Only displays outdoor snow clearing but is less specific about multiple people or SUV.\n4th=[5] Shows some snow activity, but mainly involves unrelated topics.\n5th=[1] Lacks a coherent sequence and key actions described in the query.\n</summary>\n</think>\n<answer> [3] > [2] > [4] > [5] > [1] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_BYLxSOPFOuc.mp4", "activitynet/videos/v_3J_-uEjb8UE.mp4", "activitynet/videos/v__-zOtZZ_fwI.mp4", "activitynet/videos/v_3UbQ0UDmbbw.mp4", "activitynet/videos/v_C8IEqXTBod0.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A group of fish swim around in the ocean floor. A long yellow fish passes the camera. Scuba divers gather on a boat as a man speaks. We see several fish and a shark. The scuba divers dive, looking at the fish and waving at the camera.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Features a scuba diving trip with divers underwater, exploring a wreck and the ocean floor. No specific mention or view of a yellow fish or shark.\n\n[2] Shows a scuba diving excursion with divers in the water and on a boat, encountering marine life. No prominent yellow fish or shark is visible.\n\n[3] Displays divers underwater among coral reefs and marine life, but no specific yellow fish or shark is visible.\n\n[4] Primarily focused on boat activities and snorkeling. Little focus on underwater marine life, no specific mention of yellow fish or shark.\n\n[5] Contains divers interacting with a variety of fish, including a very visible long yellow fish passing by, and a shark. Divers prepare on a boat, diving, and waving at the camera.\n</content>\n\n<contrast>\n- [5] is aligned with the query because it explicitly shows a long yellow fish, a shark, and divers interacting underwater, whereas most do not focus on marine creatures or the requested sequence.\n- [2] shows divers underwater and relates better than [1], but lacks the specific presence of a yellow fish and shark.\n- [3] features underwater exploration but misses key query elements like the yellow fish and shark.\n- [1] and [4] fall short as there is no mention of notable fish or shark interactions.\n</contrast>\n\n<summary>\n1st=[5] (features long yellow fish, shark, divers on boat and underwater)\n2nd=[2] (divers underwater, some marine life interaction)\n3rd=[3] (divers with marine life, no yellow fish or shark)\n4th=[1] (divers underwater, lacks marine life focus)\n5th=[4] (focus on boating, minimal underwater interaction)\n</summary>\n</think>\n<answer> [5] > [2] > [3] > [1] > [4] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_kuPWb9E4aUQ.mp4", "activitynet/videos/v_Zob68l3snFo.mp4", "activitynet/videos/v_fEsTL9tYOVc.mp4", "activitynet/videos/v_f9eW4rFQG-A.mp4", "activitynet/videos/v_LrwBGQ9B0Vg.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"A man is seen kneeling down next to a dog scrubbing his back while another man pours water over the dog's body. The man continues scrubbing as the other pours water and ends with the dog running away.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] The sequence depicts multiple shots of a dog being bathed with shampoo, but only one person interacting with the dog. There is no visible pouring of water by another person or action ending with the dog running away.\n\n[2] This sequence shows a man kneeling and scrubbing a dog while another man pours water over it, followed by the dog beginning to move away. This closely matches the query description.\n\n[3] Shows a small dog being bathed inside a tub by one person, lacking the element of two people involved and ending action of the dog running away.\n\n[4] Features a person washing a dog in a tub, missing the element of collaboration between two people, and no indication of the dog running away.\n\n[5] Depicts a professional dog grooming setting with one person handling the dog. The setting and actions do not match the query's outdoor, two-person scenario.\n</content>\n\n<contrast>\n- [2] matches the query as it involves two people actively bathing the dog with one scrubbing and the other pouring water, followed by the dog moving away, whereas [1] lacks the second person pouring water and does not conclude with the dog running.\n- [1] includes the washing action, but lacks the query's specifics about the pouring of water by another person or the dog running away, which [2] captures.\n- [3] is mismatched as it occurs in a small indoor setting and lacks the action involving two people.\n- [4] has only one person involved and does not depict the dog running, whereas [2] includes both requested actions.\n- [5] is far removed in its professional indoor setting and lacks two-person involvement, unlike [2]'s collaborative and outdoor setup.\n</contrast>\n\n<summary>\n1st=[2] (matches query precisely with two-person involvement and dog running away)\n2nd=[1] (related content but missing key query elements)\n3rd=[3] (includes dog bath but lacks two people and final dog action)\n4th=[4] (single person and no running dog conclusion)\n5th=[5] (indoor setting and single person, far from query)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_iAWXaWo82u4.mp4", "activitynet/videos/v_8ClpQ4q6NmM.mp4", "activitynet/videos/v_zJNYqTVWqyY.mp4", "activitynet/videos/v_17ytsjuUCRA.mp4", "activitynet/videos/v_J7LhH4IKvnM.mp4" ]
[ { "content": "System Prompt: You are RankLLM, a vision-only reranker. You will receive a Query and N candidates. Each candidate is a video described via frames. \n\nDecision rules (concise):\n- Judge relevance ONLY from what is visible in the frames.\n- Down-rank clear mismatches (wrong domain/scene/action). Consider temporal coherence.\n- Tie-breakers (in order): action visibility/close-up > temporal coherence > low occlusion/clutter > overall coverage.\n\nABSOLUTE OUTPUT CONTRACT:\n<think>\n<content>\n[1] ...\n[2] ...\n[3] ...\n[4] ...\n[5] ...\n</content>\n\n<contrast>\n...\n</contrast>\n\n<summary>\n1st=[i] ...\n2nd=[j] ...\n3rd=[k] ...\n4th=[m] ...\n5th=[n] ...\n</summary>\n</think>\n<answer> [i] > [j] > [k] > [m] > [n] </answer>\n\nConstraints:\n- Inside <think>, use ONLY <content>, <contrast>, <summary>.\n- Do NOT use per-candidate XML tags.\n- Always output a total order in <answer>.\n\n\nQuery: \"One man is explaining to another man how to use a rowing machine. One of the main starts to row on the rowing machine before the other. Next, the other man joins the first man and starts to row.\"\n\nCandidates:\n[1] video:<video>\n[2] video:<video>\n[3] video:<video>\n[4] video:<video>\n[5] video:<video>\n", "role": "user" }, { "content": "<think>\n<content>\n[1] Two men are using rowing machines side by side. Initially, one man rows alone, then they both start rowing together. This aligns with the query.\n\n[2] The video is instructional, showing one person and emphasizing learning and technique. It doesn't fit the query's specifics about two men rowing together.\n\n[3] A man is shown demonstrating on a rowing machine, possibly explaining to another person. However, only one person is rowing at a time, which partially fits the query.\n\n[4] A single individual is seen using the rowing machine without interaction with another, missing the two-person interaction in the query.\n\n[5] A large group competition setting with multiple people rowing is depicted, not matching the two-person individual interaction described in the query.\n</content>\n\n<contrast>\n- [1] beats [3] because it shows two men rowing together as the query specifies, whereas [3] only shows one person rowing at a time.\n- [3] beats [2] because it suggests some explanation may occur, whereas [2] is purely instructional with no interaction.\n- [2] ranks higher than [4] because it at least addresses the rowing machine context, whereas [4] shows individual action.\n- [4] beats [5] because it features a rowing machine being used by a man, albeit just one, whereas [5] focuses on a competition and lacks the personal instruction element.\n</contrast>\n\n<summary>\n1st=[1] (fully matches query with two men both rowing and explaining)\n2nd=[3] (partial match with some indication of explanation)\n3rd=[2] (instructional, touches on technique but lacks interaction)\n4th=[4] (one man rowing, no interaction)\n5th=[5] (group competition, not specific to query)\n</summary>\n</think>\n<answer> [1] > [3] > [2] > [4] > [5] </answer>", "role": "assistant" } ]
[ "activitynet/videos/v_o5j6N8O6HS4.mp4", "activitynet/videos/v_RpyRgnHdUXc.mp4", "activitynet/videos/v_-ceM4HHukhE.mp4", "activitynet/videos/v_Iib8Qt_9HQc.mp4", "activitynet/videos/v_FjY3eKNvzug.mp4" ]