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 explains and demonstrates how to use exercise equipment, showcasing his technique and results.\"\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 person using a NordicTrack exercise machine, focusing on its features and showing the person using the equipment. Although it demonstrates how to use the equipment, it focuses more on features and brand rather than technique or results.\n[2] The video shows a person wearing stilts or jumping shoes in an outdoor park setting, walking or engaging in a physical activity. The focus is on the activity rather than specific exercise equipment technique.\n[3] The video displays a man seated and lying down on an exercise bench, explaining techniques for using the bench. He is mostly talking and occasionally demonstrates exercises, which is relevant to explaining and demonstrating exercise equipment.\n[4] The video focuses on a man in a gym showing how to use various exercise machines, clearly demonstrating techniques and possibly discussing results, fitting the query perfectly.\n[5] The video is about a man explaining and demonstrating the use of vacuum cleaning equipment indoors, which is irrelevant to exercise equipment.\n</content>\n\n<contrast>\n- [4] ranks higher than [1] because the focus is on gym equipment specifically, with visible demonstration and potential technique explanation, whereas [1] is more brand-focused.\n- [3] ranks higher than [1] due to clearer demonstrations and verbal explanations of technique by an individual on the exercise bench.\n- [1] ranks higher than [2] because it shows exercise equipment, whereas [2] focuses on physical activity not tied to equipment use/explanation.\n- [2] ranks higher than [5] because it relates more to physical activity even if not equipment-focused, whereas [5] is unrelated to exercise.\n</contrast>\n\n<summary>\n1st=[4] (clear demonstration of exercise equipment use and technique)\n2nd=[3] (verbal explanation and demonstration of exercise techniques)\n3rd=[1] (demonstrates equipment use but focused more on brand)\n4th=[2] (shows physical activity but not focused on exercise equipment)\n5th=[5] (unrelated to exercise equipment)\n</summary>\n</think>\n<answer> [4] > [3] > [1] > [2] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_NzRNlTfyYtI.mp4",
"activitynet/videos/v_NLTd7qCLT7o.mp4",
"activitynet/videos/v_FujUrbL1qc0.mp4",
"activitynet/videos/v_AuVL_nkYXw0.mp4",
"activitynet/videos/v_C1TWr5XH8b0.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 play an exciting game of beach volleyball, with the match culminating in a victory for the team featuring a man and a woman.\"\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 involves an indoor futsal match, showing players from two opposing teams (one in red, the other in yellow) competing on a blue court with a focus on action near the goal and celebrations.\n[2] The footage displays a beach volleyball game with several women playing. It shows the game in progress, but there's no clear indication of victory or involvement of a mixed-gender team.\n[3] The frames depict a beach soccer setup with a group of people. It includes shots of a sandy court and close-ups, but lacks a focus on volleyball or a mixed-gender climax.\n[4] This sequence captures a casual game of beach volleyball near the water, played by two men. The action appears relaxed, with no definitive competitive elements visible.\n[5] The video shows a mixed-gender team playing beach volleyball. It culminates with celebratory actions, suggesting a possible victory fitting the query's requirement of a team with a man and a woman being involved in an exciting game and winning.\n\n</content>\n<contrast>\n- [5] is better than [2] because it features a mixed-gender team playing beach volleyball with clear celebratory actions suggesting a victory, whereas [2] shows only women playing without clear victory indicators.\n- [2] is preferred over [4] due to [2]'s more competitive setup and clearer beach volleyball context, while [4] appears more casual and unclear.\n- [4] beats [1] because it is at least on a beach with a volleyball net, while [1] is about indoor futsal, which is irrelevant.\n- [1] is ranked higher than [3] as futsal may be a closer sport context than beach soccer, even though both lack direct relevance.\n- [3] is the least relevant with no connection to volleyball, focusing on beach soccer content.\n\n</contrast>\n<summary>\n1st=[5] Mixed-gender team plays and seems to win beach volleyball, fulfilling the query.\n2nd=[2] Women playing beach volleyball, lacks mixed-gender element and clear victory.\n3rd=[4] Casual beach volleyball between men, lacks mixed gender and competitive climax.\n4th=[1] Indoor futsal with no volleyball content, irrelevant sport.\n5th=[3] Beach soccer, irrelevant to volleyball query context.\n\n</summary>\n</think>\n<answer> [5] > [2] > [4] > [1] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_XxMzH75Bhr8.mp4",
"activitynet/videos/v_FC5FsugGt8o.mp4",
"activitynet/videos/v_XD3yFrJHiv8.mp4",
"activitynet/videos/v_ymEh5T11kSQ.mp4",
"activitynet/videos/v_SLPD1U5iDjY.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 people take turns performing impressive kicks and tricks in the center of a circle while an enthusiastic crowd watches and reacts from the sidelines.\"\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 group of people in orange shirts are on a field, possibly engaged in a sports activity, but there are no visible kicks or tricks. The focus is more on people standing or walking, with a calm setting and spectators in the background.\n\n[2] People in costumes are on an indoor sports field, participating in what appears to be a humorous or entertaining event. The crowd is lively, but the activities do not clearly center on kicking or tricks.\n\n[3] A sumo wrestling match is taking place with a large audience watching. The action involves wrestling rather than kicking or tricks, and it does not fit the query of kicks and tricks in a circle.\n\n[4] Several individuals are performing martial arts or dance-like movements in a circle, surrounded by a crowd. The focus is on physical tricks and kicks performed in the center, while the audience watches and reacts enthusiastically.\n\n[5] Rodeo activities are shown, with horse riding and cattle events. The focus here is on rodeo-style competitions, not on kicks or tricks performed by people in a circle.\n</content>\n\n<contrast>\n- [4] is the best match as it clearly shows people performing kicks and tricks within a circle, with an engaged crowd responding from the sidelines.\n- [1] is less relevant as it shows a calm sports-like setting without kicking or tricks.\n- [2] involves a performance with a crowd but lacks the specific kicking or trick actions.\n- [3] is focused on wrestling, which does not match the described activity of performing tricks in a circle.\n- [5] displays rodeo activities, which do not involve the type of kicks and tricks described in the query.\n</contrast>\n\n<summary>\n1st=[4] (People perform tricks and kicks in a circle with audience engagement)\n2nd=[2] (Costumed event with a crowd, lacks the specific kicking/tricks)\n3rd=[1] (Sports-like setting without kicks and tricks)\n4th=[3] (Wrestling action, not trick performance)\n5th=[5] (Rodeo activities focus, not on kicks/tricks)\n</summary>\n</think>\n<answer> [4] > [2] > [1] > [3] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_0jBwj0bfZ3Y.mp4",
"activitynet/videos/v_hK3E6njydQA.mp4",
"activitynet/videos/v_slUvIngXl4g.mp4",
"activitynet/videos/v_AkeJ-lJLNkw.mp4",
"activitynet/videos/v_w9cJ3FNGw_U.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 work together to move and spread mulch and dirt in a yard using shovels, rakes, a wheelbarrow, and a small tractor.\"\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 chopping wood using an axe in a yard, with wood pieces and a wheelbarrow nearby. There is no evidence of mulch, dirt, or group activity with different tools.\n[2] Children are playing outdoors, hitting a hanging piñata. This is unrelated to the query about yard work with tools.\n[3] A toddler is playing with a lacrosse stick on a wooden deck. There is no evidence of yard work or people working with tools.\n[4] People are working in a yard spreading mulch with rakes, shovels, a wheelbarrow, and a small tractor visible. This closely matches the query description of collaborative yard work.\n[5] People are chopping wood with axes and carrying logs. There is no evidence of spreading mulch or dirt with multiple tools.\n</content>\n\n<contrast>\n- [4] contains multiple people and tools used in spreading mulch, aligning with the yard work context, whereas [1] and [5] involve wood chopping.\n- [2] features children playing with a piñata, unrelated to yard work, while [4] shows active yard maintenance.\n- [3] is unrelated, showing a toddler playing rather than any yard work activities, while [4] involves the specified yard work activities.\n</contrast>\n\n<summary>\n1st=[4] (Matches the query with people spreading mulch using shovels, rakes, a wheelbarrow, and a small tractor)\n2nd=[1] (Involves yard work but limited to wood chopping and lacks group and tool diversity)\n3rd=[5] (Similar yard setting but focused on wood chopping, not mulch spreading)\n4th=[3] (Contains a child playing with a stick, unrelated to the query)\n5th=[2] (Involves a piñata, unrelated to yard work)\n</summary>\n</think>\n<answer> [4] > [1] > [5] > [3] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_b2uAgV5n3FU.mp4",
"activitynet/videos/v_3HP2E_a9xrU.mp4",
"activitynet/videos/v_CG-7jcSB5_c.mp4",
"activitynet/videos/v_FhWxgGQtCJY.mp4",
"activitynet/videos/v_yiFKPKXevOU.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 boys play soccer and pass a ball to each other on a busy beach as people walk by and enjoy the water.\"\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 children swimming and playing on a beach but does not feature soccer or a ball being passed.\n[2] This video depicts two boys on a beach passing a ball to each other, fitting the query description well. There is visible beach activity in the background.\n[3] The video shows a professional beach soccer match setting with organized teams and a crowd, which does not align with the recreational and informal query requirement.\n[4] The video focuses mainly on beach volleyball with no evidence of soccer or two boys playing individually. The scene appears crowded with beachgoers.\n[5] The video shows a beach volleyball game with no sign of a soccer ball or two boys playing soccer. It captures a casual beach environment with some activity in the background.\n</content>\n\n<contrast>\n- [2] fits the query because it shows two boys playing with a ball on a beach, whereas [1] shows no ballplay.\n- [2] is more relevant than [3] because it features casual play instead of a professional setting.\n- [2] is superior to [4] and [5] which focus on volleyball and lack the described soccer activity.\n</contrast>\n\n<summary>\n1st=[2] (features two boys passing a ball on a beach, fits query well)\n2nd=[1] (shows children on a beach, but no soccer activity)\n3rd=[3] (professional beach soccer but not the informal play described)\n4th=[4] (features beach volleyball, not soccer)\n5th=[5] (focused on volleyball, lacks soccer element)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [4] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_tQmsDeu1d6M.mp4",
"activitynet/videos/v_ymEh5T11kSQ.mp4",
"activitynet/videos/v_09MaNbzc2TA.mp4",
"activitynet/videos/v_KNpeiKm3xyM.mp4",
"activitynet/videos/v_SLPD1U5iDjY.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 stands on a roof, speaks briefly, and proceeds to repair the roof by removing, organizing, and replacing tiles and wooden parts.\"\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 interior work, involving people in a room laying floor tiles. There are several sequences displaying the process of placing and setting these tiles. The actions are tile-centric, not related to roofing.\n\n[2] This video shows a man on a roof, as per the query. It starts with him standing and presumably speaking, followed by the removal and organization of roof tiles. The man is involved in direct roof repair work, aligning closely with the query.\n\n[3] The video captures a sequence of a person working on a roof installing a vent pipe. There's a focus on sealing and adjusting the vent roof component. However, it doesn't involve organizing or replacing tiles or parts.\n\n[4] This video displays a person performing maintenance on a sloped roof. The person appears to be checking or fixing something on the roof's edge, but the specific actions of removing, organizing, or replacing tiles or wooden parts are not clear.\n\n[5] The video shows a man applying plaster or paint on a wall indoors. The actions are related to wall finishing and have no association with roofing activities or the sequence requested in the query.\n</content>\n\n<contrast>\n- [2] is preferred over [3] because it clearly shows roof repair involving tile removal and organization, which matches the query best, whereas [3] focuses on a vent installation without tile work.\n- [2] is preferred over [4] because [2] demonstrates clear actions of tile removal and reorganization, whereas [4] shows general roof maintenance without clarity on tile work.\n- [4] beats [3] as the actions occur on a roof, offering some relation, though indirect, to the query, whereas [3] is more specialized on vent installation.\n- [1] and [5] are ranked lower because they involve activities completely unrelated to roofing (tile flooring and wall plastering, respectively).\n</contrast>\n\n<summary>\n1st=[2] Demonstrates roof work with tile removal and replacement.\n2nd=[4] Shows a man on a roof with some maintenance activity.\n3rd=[3] Involves roof work but focuses on vent pipe adjustments.\n4th=[1] Entirely focused on indoor tile flooring work.\n5th=[5] Displays wall plastering, unrelated to roofing.\n</summary>\n</think>\n<answer> [2] > [4] > [3] > [1] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_fVg3StD8LL0.mp4",
"activitynet/videos/v_6sorAV7KJxA.mp4",
"activitynet/videos/v_mCdKLZGnKxk.mp4",
"activitynet/videos/v_fljxcvQe_Mo.mp4",
"activitynet/videos/v_MZ4yIxM6snw.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 enjoys an exhilarating water skiing experience, skillfully maneuvering across the water while others watch from a nearby boat.\"\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 water skiing skillfully on a lake, with a noticeable wake trail. Occasionally, other people appear on a boat watching the activity. This directly matches the query of a man water skiing while being observed.\n\n[2] This sequence features a mix of activities, primarily water skiing. There are dynamic shots of a person water skiing with impressive maneuvers, but there is no visible interaction with spectators from a boat, which is a part of the query.\n\n[3] The video depicts whitewater kayaking through rapids, with scenes of individuals navigating the water. This does not relate to water skiing or the presence of observers on a boat.\n\n[4] This sequence is about surfing, where a person rides waves on a surfboard. It does not include water skiing or boat watchers, making it irrelevant to the query.\n\n[5] The video consists of snow tubing on a slope, showing people enjoying sliding down the snow. There is no water skiing or presence of spectators on a boat, making it unrelated to the query.\n</content>\n<contrast>\n- [1] is ranked higher than [2] because it showcases water skiing with spectators from a boat, aligning perfectly with the query, whereas [2] lacks the spectator aspect.\n- [2] is ranked higher than [4] as it features water skiing, though lacks boat spectators, which is closer to the query than surfing.\n- [4] is ranked higher than [3] because surfing is more related to water activities than kayaking through rapids, which do not fit the query.\n- [3] is ranked higher than [5] because whitewater kayaking is more closely related to aquatic sports than snow tubing, which is not related to the query.\n</contrast>\n<summary>\n1st=[1] (full match to query: water skiing with spectators on a boat)\n2nd=[2] (water skiing without boat spectators)\n3rd=[4] (related as an aquatic sport: surfing)\n4th=[3] (aquatic sport: kayaking, less relevant)\n5th=[5] (unrelated activity: snow tubing)\n</summary>\n</think>\n<answer> [1] > [2] > [4] > [3] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_-5KAycAQlC4.mp4",
"activitynet/videos/v_G3TPqsLNlrw.mp4",
"activitynet/videos/v__3lgJ6yp7o8.mp4",
"activitynet/videos/v_9ZGy24Geu5k.mp4",
"activitynet/videos/v_6TgVNz6PATk.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 demonstrates how to change a tire tube by removing the old one, inserting a new tube, and inflating it.\"\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 depict a snowy environment where a person is clearing snow from a car, indicative of weather-related maintenance but unrelated to tire tube changing.\n[2] This series shows a person starting a fire using sticks and paper, an action unrelated to changing a tire tube.\n[3] These images display individuals working on a large vehicle's wheel, possibly removing and replacing parts. It appears relevant to tire maintenance but not specific to tube changing.\n[4] Shows a person using stilts in a park setting, unrelated to tire tube changing.\n[5] Clearly demonstrates a man changing a bicycle tire tube, including removing the old tube, inserting a new one, and inflating it, matching the query exactly.\n</content>\n\n<contrast>\n- [5] is ranked highest because it explicitly shows step-by-step tire tube changing.\n- [3] displays wheel-related work but lacks clear tube replacement, whereas [5] does.\n- [1] involves automotive maintenance but not tire-related, whereas [5] does.\n- [2] and [4] are unrelated to tire tube changing; therefore, [3] is prioritized over these despite lesser accuracy than [5].\n- [2] and [4], both unrelated, show outdoor activities with [2] involving fire and [4] stilt walking, which do not involve any wheel or tire activity.\n</contrast>\n\n<summary>\n1st=[5] (Explicitly matches tire tube changing with clear actions)\n2nd=[3] (Involves wheel-related repair but not as clear on tube change)\n3rd=[1] (Involves car maintenance but unrelated to tube changing)\n4th=[2] (Fire lighting activity, unrelated)\n5th=[4] (Stilt walking activity, unrelated)\n</summary>\n</think>\n<answer> [5] > [3] > [1] > [2] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_3UbQ0UDmbbw.mp4",
"activitynet/videos/v_K-kJ-0nkKZk.mp4",
"activitynet/videos/v_YoJxVl_hUWg.mp4",
"activitynet/videos/v_NLTd7qCLT7o.mp4",
"activitynet/videos/v_C_bwHYiX-Vw.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 showcases various athletes participating in shotput competitions, taking turns to throw while spectators observe from the sidelines.\"\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 athletes participating in javelin competitions. The frames display competitors taking turns to throw the javelin, with distances marked and spectators in the stands.\n[2] The video captures a group of individuals standing on a field, possibly engaging in casual or practice activities. The setup doesn’t suggest any competitive sporting event.\n[3] This video aligns with the query. It shows athletes in an indoor arena participating specifically in shotput competitions. Several athletes take turns, and measures or judges are visible in the background.\n[4] The video is about a men’s triple jump competition. It focuses on athletes performing jumps, with event details displayed prominently.\n[5] The video depicts a historical athletic event focusing on a triple jump competition. Athletes are seen preparing and executing jumps with spectators in the stadium.\n\n</content>\n\n<contrast>\n- [3] features shotput events clearly and matches the query, whereas [1] showcases javelin, a different event.\n- [3] displays shotput actions indoors in a competition, whereas [2] shows casual field activity with no clear shotput event.\n- [3] directly involves shotput with detailed action sequences, whereas [4] focuses on triple jump, unrelated to shotput.\n- [3] centers on shotput competitions; [5] is focused on triple jump, which differs from the query topic.\n\n</contrast>\n\n<summary>\n1st=[3] (Includes shotput events with athletes competing and detailed sequences that match the query)\n2nd=[1] (Displays athletes but in javelin events, not shotput)\n3rd=[4] (Covers a triple jump competition, unrelated but with a competition setup)\n4th=[5] (Similarly engages in a triple jump rather than shotput)\n5th=[2] (Shows a generic field scene with no competitive event indication)\n\n</summary>\n</think>\n<answer> [3] > [1] > [4] > [5] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_sO7ZGxFwbm8.mp4",
"activitynet/videos/v_0jBwj0bfZ3Y.mp4",
"activitynet/videos/v_hniQpwn3Ob4.mp4",
"activitynet/videos/v_cwQQj3VZLC8.mp4",
"activitynet/videos/v_tnXjdv_GM9g.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 prepare for and enjoy a snowboarding adventure in a snowy mountain range, traveling by helicopter to the mountaintop and capturing their experience on 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] This sequence is about BMX biking indoors on a dirt track with riders competing, showcasing jumps and stunts. It doesn't relate to the snowboarding adventure theme.\n\n[2] The video depicts a winter sports event with snowboarding and skiing at a snowy location. There are competitors and spectators, but no visible helicopter or mountain travel.\n\n[3] This sequence shows a person painting various designs, unrelated to snowboarding or outdoor adventure.\n\n[4] The video captures a person waterskiing behind a boat, unrelated to a snowy mountain adventure or snowboarding.\n\n[5] This sequence shows a snowy mountain range with people preparing and possibly engaging in snow sports. It includes wide shots of the mountains and individuals in snowboarding gear, fitting the query well.\n\n</content>\n\n<contrast>\n- [5] includes snowy mountains, preparing individuals, and scenes coherent with the snowboarding adventure, whereas [2] lacks travel by helicopter and full preparation context.\n- [2] features winter sports but shows more of a competitive event atmosphere, whereas [5] aligns more with the leisurely, adventurous essence of the query.\n- [1] involves BMX biking which is unrelated to snowboarding, lacking the snow and mountain environment.\n- [3] and [4] are completely off-topic, focusing on painting and waterskiing respectively.\n- [5] beats [2] because it better encompasses the adventurous preparation aspect of the query.\n- [2] has snow sports visuals, but with a sports competition focus, unlike [5].\n- [4] captures water-based action, standing far from the snowy mountain theme.\n- [3] provides an indoor painting process, irrelevant to outdoor snowboarding.\n \n</contrast>\n\n<summary>\n1st=[5] (Best visual alignment with snowy mountain adventure and preparation)\n2nd=[2] (Shows snow sports but leans towards competition rather than adventure)\n3rd=[1] (BMX biking, unrelated to snowboarding)\n4th=[4] (Waterskiing, unrelated to the query's theme)\n5th=[3] (Painting, completely off-topic from snowboarding adventure)\n</summary>\n</think>\n<answer> [5] > [2] > [1] > [4] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Vfb3LmFDols.mp4",
"activitynet/videos/v_FNB1BMEtIvw.mp4",
"activitynet/videos/v_N_zbcVs419E.mp4",
"activitynet/videos/v_G5mKcPzVMG4.mp4",
"activitynet/videos/v_Lr0Vm9zBXKE.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 and woman host a talk show from a studio, discussing and watching video clips featuring men playing outdoor soccer in a city setting.\"\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 and woman hosting a talk show in a studio setting. They are sitting on a couch, discussing sports-related topics. The video also shows clips of men playing soccer on a sandy outdoor city setting, which aligns closely with the query description.\n\n[2] The video captures two men in a backyard setting playing croquet. This does not match the query about hosting a talk show and soccer in a city setting.\n\n[3] The video involves a man and a woman discussing car-related topics within a studio and outdoor environments. It involves cars and driving, which do not relate to the query focused on soccer and talk shows.\n\n[4] This video features a cooking show with a chef and a woman demonstrating cooking techniques in a kitchen. It is unrelated to the query about soccer or a talk show.\n\n[5] The video presents a tutorial or entertainment segment on \"How to Win Rock, Paper, Scissors.\" It features a man and woman, but they aren't hosting a talk show or discussing soccer.\n</content>\n\n<contrast>\n- [1] is most relevant because it features a talk show setting and includes video clips of outdoor soccer, matching the query directly, whereas [2], [3], [4], and [5] do not involve soccer or talk show discussions of soccer.\n- [2] focuses on croquet in a backyard, whereas [1] has a clear focus on soccer clips relevant to the query.\n- [3] involves car discussions and activities, whereas [1]'s soccer theme aligns with the query.\n- [4] is about cooking, with no relationship to soccer or a talk show format, unlike [1].\n- [5] shows a game tutorial, lacking any soccer or talk show elements that [1] includes.\n</contrast>\n\n<summary>\n1st=[1] (Direct match with query: talk show and soccer clips.)\n2nd=[3] (Studio setup discussion, but unrelated topic.)\n3rd=[5] (Interaction of a man and woman with a minor talk show element.)\n4th=[2] (Features an outdoor game but not soccer or a city setting.)\n5th=[4] (Least relevant; focused on cooking, no link to soccer or talk shows.)\n</summary>\n</think>\n<answer> [1] > [3] > [5] > [2] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_XD3yFrJHiv8.mp4",
"activitynet/videos/v_eypKc4FyZ2s.mp4",
"activitynet/videos/v_PZjWkB_q2lE.mp4",
"activitynet/videos/v_xZdCmghMcvU.mp4",
"activitynet/videos/v_AnRYS02tvRA.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 demonstrates the step-by-step process of fixing a car tire onto a wheel rim, including the necessary tools and techniques for assembly, tightening, and checking tire pressure.\"\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 appears to show a man working on a bicycle wheel, possibly assembling the tire. Tools and steps involving a bicycle tire are visible, suggesting a demonstration on this topic. However, it focuses on a bicycle, not a car.\n\n[2] The sequence shows a man on a workout bench, possibly demonstrating an exercise routine. This content is unrelated to fixing a car tire.\n\n[3] The video depicts a man in a garage setting, with cars and tires visible. There appears to be a discussion on tire-related topics, but not a detailed step-by-step tire mounting process.\n\n[4] This sequence shows a man working with a car tire and rim. A variety of tools are displayed, and there are steps showing the tire being mounted on the rim, tightening, and adjustments, aligning closely with the query.\n\n[5] The video depicts people working on a vehicle wheel, involving tools and some steps are unclear due to the outdoor setting and less focus on the tire-rim assembly. Though it involves a tire, it lacks the clarity and step-by-step demonstration shown in [4].\n</content>\n\n<contrast>\n- [4] closely matches the query with clear step-by-step instructions using tools on a car tire, whereas [3] lacks specific assembly details.\n- [3] has relevant car and tire content, but focuses more on tire information rather than hands-on assembly.\n- [5] involves a car tire but lacks clarity in demonstrating specific techniques or a systematic process as seen in [4].\n- [1] shows a detailed step-by-step process but is about a bicycle tire, not a car, whereas [4] directly addresses car tires.\n- [2] is unrelated as it involves exercise demonstration, not tire fixing.\n</contrast>\n\n<summary>\n1st=[4] (detailed step-by-step process with tools on a car tire)\n2nd=[3] (relevant car tire information, lacks assembly specifics)\n3rd=[5] (some related activity with a car tire, less clear process)\n4th=[1] (detailed process but focuses on bicycle tire)\n5th=[2] (unrelated to tire fixing, focuses on exercises)\n</summary>\n</think>\n<answer> [4] > [3] > [5] > [1] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_C_bwHYiX-Vw.mp4",
"activitynet/videos/v_FujUrbL1qc0.mp4",
"activitynet/videos/v_9ddD5Ob93J0.mp4",
"activitynet/videos/v_4r0qYLCucEI.mp4",
"activitynet/videos/v_YoJxVl_hUWg.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: \"Scuba divers explore the ocean, observing various fish and a shark while interacting with the camera both on a boat and underwater.\"\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 windsurfing activities. The sequence includes people windsurfing on the ocean and interacting with the camera onboard a boat. It lacks any underwater scenes or interaction with marine life typical for scuba diving.\n\n[2] This video includes sequences of scuba diving with scenes underwater featuring fish and divers. There are views from a boat and interaction underwater, aligning with the query of scuba divers exploring the ocean and observing marine life.\n\n[3] This video primarily features scuba diving and underwater exploration. Divers can be seen entering the water from a boat, and there are scenes of marine life, including jellyfish and other underwater elements. However, there is no interaction with a shark visible.\n\n[4] The video is a scuba diving adventure, with scenes of divers underwater observing various fish, including a shark. There are both above-water and underwater interactions which closely match the query.\n\n[5] This video takes place in a swimming pool with people practicing swimming techniques. There is no ocean exploration or marine life interaction.\n\n</content>\n\n<contrast>\n- [4] aligns strongly with the query because it involves scuba divers, marine life including a shark, and both boat and underwater interactions; whereas [2] lacks a clear shark interaction.\n- [2] is relevant with its underwater exploration and fish observation, whereas [3] lacks the shark but shows divers entering from a boat.\n- [3] involves scuba divers and underwater scenes, but there is no clear interaction with a shark, unlike [4].\n- [1] involves windsurfing, completely unrelated to scuba diving or marine life exploration.\n- [5] occurs in a swimming pool, unrelated to oceanic or marine life exploration.\n\n</contrast>\n\n<summary>\n1st=[4] (includes scuba diving, marine life, shark presence, boat/underwater interaction)\n2nd=[2] (scuba diving and marine life exploration, no shark)\n3rd=[3] (scuba diving and underwater, no clear shark)\n4th=[1] (windsurfing, unrelated to marine life exploration)\n5th=[5] (pool swimming, unrelated to ocean exploration)\n\n</summary>\n</think>\n<answer> [4] > [2] > [3] > [1] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_AhWkUdag6NA.mp4",
"activitynet/videos/v_PT18KK_6jmU.mp4",
"activitynet/videos/v_hpU50i4WBK8.mp4",
"activitynet/videos/v_LrwBGQ9B0Vg.mp4",
"activitynet/videos/v_XC6tvSBS0PA.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 boy performs an energetic drum solo, playing both the drums and cymbals.\"\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 in traditional attire performing with drums. There is no young boy visible. The setting is outdoors with traditional elements, but it does not match the query specifics.\n\n[2] The video features a person, not a young boy, playing a drum set indoors. The focus is on the drums, but the performer does not match the age specified in the query.\n\n[3] This video includes multiple performers, but none appear to be a young boy. There is a focus on drums, but the context involves additional instruments, which does not match the query closely.\n\n[4] The video depicts a foosball tournament, unrelated to a drum performance. No musical instruments or young boys are visible, making it irrelevant to the query.\n\n[5] The video shows a young boy energetically playing a drum kit, including cymbals. The setting is indoors, with close-ups on the performance. It is highly relevant to the query, as it captures the action and specifics requested.\n</content>\n\n<contrast>\n- [5] is highly relevant because it depicts a young boy playing drums and cymbals energetically, matching the query directly.\n- [2] shows a drum performance but by an adult, whereas [5] features a young performer as specified.\n- [3] includes drums but involves multiple performers and instruments, lacking the focus on a young boy.\n- [1] involves drumming but not by a young boy, whereas [5] matches the age and action specifics.\n- [4] is unrelated, focusing on a foosball game, whereas [5] perfectly depicts the query.\n</contrast>\n\n<summary>\n1st=[5] (perfect match: young boy energetically playing drums and cymbals)\n2nd=[2] (involves drums but not a young boy)\n3rd=[3] (features drums but lacks focus on a single young performer)\n4th=[1] (involves drumming but not by a young boy)\n5th=[4] (unrelated content)\n</summary>\n</think>\n<answer> [5] > [2] > [3] > [1] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_GHcNgllmcpM.mp4",
"activitynet/videos/v_DYahQgHG1Gk.mp4",
"activitynet/videos/v_zm8gZpwZkRE.mp4",
"activitynet/videos/v_iC1C5wfb0iU.mp4",
"activitynet/videos/v_FRsHew2Pcko.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 demonstrates lifting, carrying, and lunging with weights in a gym setting.\"\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 using an elliptical machine. The sequence shows exercise with the machine, without any lifting, carrying, or lunging with weights.\n[2] The person demonstrates lifting a barbell and performing a lunge. It includes preparatory movements, lifting, and balanced lunging in a gym setting.\n[3] The sequence shows a person lifting a barbell. There is no clear demonstration of carrying or lunging.\n[4] The person is performing a snatch, showcasing weightlifting. Lifting is visible, but not the full range of carrying and lunging.\n[5] The sequence shows a dance performance in a gymnasium setting. It involves typical dance movements without weights.\n</content>\n\n<contrast>\n- [2] is ranked higher because it includes lifting, carrying, and lunging with weights, whereas [3] only shows lifting without any lunging or carrying.\n- [3] demonstrates weightlifting, but lacks the carrying and lunging present in [2].\n- [4] includes lifting, similar to [3], but lacks additional elements like carrying and lunging seen in [2].\n- [1] and [5] do not show any lifting, carrying, and lunging with weights, focusing on different activities instead.\n</contrast>\n\n<summary>\n1st=[2] (demonstrates lifting, carrying, and lunging with weights)\n2nd=[3] (focuses on weightlifting but lacks carrying and lunging)\n3rd=[4] (shows lifting but lacks carrying and lunging)\n4th=[1] (exercises on an elliptical machine, not related to query)\n5th=[5] (a dance routine, unrelated to weightlifting activities)\n</summary>\n</think>\n<answer> [2] > [3] > [4] > [1] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_3cmHM1Yu92s.mp4",
"activitynet/videos/v_3oJ62-Ce9ys.mp4",
"activitynet/videos/v_u08gU4eQFHY.mp4",
"activitynet/videos/v_xMQcYD044z4.mp4",
"activitynet/videos/v_8w9pf_Ecm_U.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 and his wife float down a river on intertubes with a large group, enjoying the scenery and camaraderie along the way.\"\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 a group of people riding horses through a natural setting, including river crossings. There are no intertubes or floating involved, which mismatches the query.\n\n[2] The video shows people floating down a river on intertubes, wearing helmets and enjoying the activity. Some frames show multiple intertubes, aligning well with the query.\n\n[3] This video depicts a group of people white-water rafting in inflatable rafts, not on intertubes. The focus is on more intense waters, which is not a perfect match for the query's leisurely theme.\n\n[4] People are paddling in canoes down a river, engaging in playful activities on the water. While it involves a group on a river, the equipment doesn't match intertubing.\n\n[5] The video shows a group of people floating down a river on intertubes, showing scenes of relaxing, social interactions, and scenic views. This closely matches the query.\n</content>\n\n<contrast>\n- [5] is ranked above [2] because it shows people relaxing and enjoying the intertubing with a large group, directly aligning with the query's emphasis on camaraderie.\n- [2] outranks [4] as it shows intertubing rather than canoeing, which better matches the query's specification.\n- [4] beats [3] since it includes a leisurely river trip with multiple people, more closely resembling the query than the intense rafting in [3].\n- [3] outranks [1] as it involves a river activity, whereas [1] is entirely about horseback riding.\n</contrast>\n\n<summary>\n1st=[5] (closest match with group intertubing on a river)\n2nd=[2] (intertubing activity, less focus on camaraderie)\n3rd=[4] (river journey, but with canoes)\n4th=[3] (river activity, although more intense than intertubing)\n5th=[1] (no river floating; horseback riding instead)\n</summary>\n</think>\n<answer> [5] > [2] > [4] > [3] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Vq2mQvMp_c4.mp4",
"activitynet/videos/v_b_v5uGSheEw.mp4",
"activitynet/videos/v_WxefBZ7Iaow.mp4",
"activitynet/videos/v_p4YTDxTASBI.mp4",
"activitynet/videos/v_OYPyiUC4h68.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 festive parade features groups in green uniforms marching down the street while waving flags and playing various brass and percussion instruments.\"\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 depicts a bull being led down a street by several individuals. There is no sign of a festive parade or green uniforms. The focus is on the bull event.\n\n[2] This video shows a pool tournament with children playing billiards. There are no signs of a parade, instruments, or green uniforms. The setting is indoors and involves a game event.\n\n[3] The scene involves individuals riding camels and participating in camel rides at a festival or event. There are no parades or groups in uniform depicted.\n\n[4] This video shows a marching band performance with individuals in uniforms, but the uniforms are not green. They are playing instruments, which aligns partially with the query.\n\n[5] This video features individuals in green uniforms marching down the street. They are waving flags and playing brass and percussion instruments, matching the query description of a festive parade.\n</content>\n\n<contrast>\n- [5] matches the query completely with green uniforms, flags, and instruments, whereas [4] lacks the color detail specifically mentioned.\n- [4] includes a marching band but doesn't show the green uniforms specified in the query description, whereas [5] does.\n- [1], [2], and [3] do not depict parades or match any elements of the query as [5] does with its clear depiction of a parade in green uniforms.\n</contrast>\n\n<summary>\n1st=[5] (features a parade with green uniforms, flags, and instruments)\n2nd=[4] (features a marching band with instruments, partial relevance)\n3rd=[3] (festival environment but lacks parade and uniform details)\n4th=[1] (street event but unrelated to the parade concept)\n5th=[2] (billiards tournament, completely unrelated)\n</summary>\n</think>\n<answer> [5] > [4] > [3] > [1] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_esQnfUcBbPE.mp4",
"activitynet/videos/v_SnBTp7iV2kE.mp4",
"activitynet/videos/v_SHfFrzy_fKM.mp4",
"activitynet/videos/v_xC5RVs9mXyM.mp4",
"activitynet/videos/v_jkn6uvmqwh4.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 practices serving tennis balls on a court with and without other players present.\"\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] Images show a racquetball game indoors, with players using racquets in a closed court. The focus is on gameplay, not tennis serving.\n[2] Similar to [1], this sequence involves a racquetball game with players in an enclosed space. Activity isn't related to serving tennis balls.\n[3] Features table tennis, with players at a ping pong table. It does not show serving or a full-sized tennis court.\n[4] Displays a badminton game in an indoor court. Actions resemble rallies rather than serving tennis balls.\n[5] A man is seen practicing serves on a tennis court, both alone and in the presence of another person, clearly matching the query.\n\n</content>\n\n<contrast>\n- [5] accurately shows a man practicing tennis serves with and without other players, whereas [1] and [2] show racquetball.\n- [5] has a clear depiction of tennis serving, whereas [3] shows table tennis, which does not match the query.\n- [5] uniquely displays a tennis serving scenario, whereas [4] focuses on badminton rallies, not related to tennis at all.\n- [1] and [2] are similar in domain but none like [5] which actually covers tennis serving with and without other players.\n</contrast>\n\n<summary>\n1st=[5] (Shows tennis serves being practiced clearly and matches the query)\n2nd=[2] (Shows a relevant indoor racquet sport but not tennis)\n3rd=[1] (Similar indoor racquet sport; lacks tennis-serving action)\n4th=[3] (Shows table tennis, unrelated to the query's tennis focus)\n5th=[4] (Focuses on badminton, not relevant to tennis serving)\n</summary>\n</think>\n<answer> [5] > [2] > [1] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_mhqfGS59FG0.mp4",
"activitynet/videos/v_rfxXEgc9RTQ.mp4",
"activitynet/videos/v_UPZm0tx77Vw.mp4",
"activitynet/videos/v_F7V-FqgG3T8.mp4",
"activitynet/videos/v_BryW4niZT9Y.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 performs a song on the flute, pausing briefly before finishing with a smile 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] A young girl with a blue wig is playing the violin in a room filled with bookshelves. She appears focused on playing throughout the sequence. There is no flute, pause, or smile implied.\n\n[2] A woman is seen using a hookah in a cafe setting. She exhales smoke and smiles at the camera. There is no flute or music performance depicted.\n\n[3] A woman plays a flute while sitting on a couch with orange cushions in the background. The video shows her performing, pausing, and then smiling at the camera, which matches the query.\n\n[4] A woman is smoking a cigarette in a dimly lit room. She poses and occasionally smiles. There is no flute or musical activity present.\n\n[5] A woman performs on a flute with sheet music and a Christmas tree in the background. She appears to pause and ends with a smile towards the camera, consistent with the query.\n</content>\n\n<contrast>\n- [3] is the most fitting as it shows a woman playing the flute, pausing, and smiling at the camera, aligning perfectly with the query.\n- [5] also matches the query by depicting a woman playing the flute with a pause and smile but has less close-up.\n- [1] features an instrument performance but it is a violin, not related to the flute.\n- [2] involves a hookah activity, irrelevant to musical performance.\n- [4] is centered on smoking and lacks any musical elements.\n</contrast>\n\n<summary>\n1st=[3] (perfect match with flute performance, pause, and smile)\n2nd=[5] (flute performance with pause and smile, but less focus on the face)\n3rd=[1] (instrument performance, but it's a violin, not a flute)\n4th=[2] (hookah use and smile, unrelated to query)\n5th=[4] (smoking scene, unrelated to musical performance)\n</summary>\n</think>\n<answer> [3] > [5] > [1] > [2] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_KYl67H9-4TA.mp4",
"activitynet/videos/v_svSM-UqjNWE.mp4",
"activitynet/videos/v_6HyNydVIji4.mp4",
"activitynet/videos/v_zuNQFkkyBNo.mp4",
"activitynet/videos/v_J6ScF5n_Cug.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 participates in a group bike ride around a track, capturing the experience with a GoPro 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 shows people running on a trail in a natural setting, possibly for a trail running event. Most images show a first-person perspective, but no biking or GoPro usage is apparent.\n\n[2] This sequence depicts several individuals using stilts or jumping devices in an urban area. There is no indication of biking or use of a GoPro camera.\n\n[3] This sequence features a person using pogo sticks or similar jumping devices in a park setting. There is no biking involved or use of a GoPro.\n\n[4] The frames capture a person using a skateboard in a suburban neighborhood. No biking or clear use of a GoPro camera is observed.\n\n[5] This sequence is a first-person perspective of a group bike ride on a track, clearly captured with a GoPro camera mounted on a helmet. The images display various angles and riders in action.\n\n</content>\n\n<contrast>\n- [5] is ranked highest because it clearly depicts a group bike ride on a track using a GoPro, matching the query perfectly.\n- [1] involves running in nature, not biking or using a GoPro, whereas [5] directly relates to the query.\n- [2], [3], and [4] show unrelated activities involving stilts and skateboarding, not involving bikes or GoPros.\n- [1] is slightly more relevant than [2], [3], and [4] because it involves outdoor physical activity, but still does not match as well as [5].\n- [2] and [3] are similar in their focus on jumping devices, providing no relevance to biking or GoPro use.\n- [4] involves skateboarding, which is closer to biking but still does not involve group rides or GoPros.\n\n</contrast>\n\n<summary>\n1st=[5] (perfect match with group bike ride and GoPro use)\n2nd=[1] (outdoor physical activity but no biking or GoPro)\n3rd=[4] (skateboarding, which is closer to biking but no GoPro)\n4th=[2] (stilts, unrelated to biking or GoPro use)\n5th=[3] (pogo sticks, unrelated and no biking or GoPro)\n\n</summary>\n</think>\n<answer> [5] > [1] > [4] > [2] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_PntiXZ692xM.mp4",
"activitynet/videos/v_qxlVkDpdyOA.mp4",
"activitynet/videos/v_NLTd7qCLT7o.mp4",
"activitynet/videos/v_DXIHbnyfNGA.mp4",
"activitynet/videos/v_XPN_PSadJ_A.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 boy plays hopscotch, eagerly running, jumping, and hopping across the course.\"\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 playing hopscotch. He is seen actively running, jumping, and hopping across a hopscotch course drawn on the ground.\n[2] The video captures a beach soccer game with adult players. There is no relevance to hopscotch.\n[3] This video depicts a young child playing with a lacrosse stick, not playing hopscotch.\n[4] The video shows an outdoor gathering with people engaging in different activities, including playing with a racket. It doesn’t involve hopscotch.\n[5] The video features elderly individuals playing croquet on a lawn, unrelated to hopscotch.\n</content>\n\n<contrast>\n- [1] is directly related to a young boy playing hopscotch, whereas [2] features a beach soccer match.\n- [1] shows a relevant hopscotch activity, while [3] depicts a child playing with a lacrosse stick.\n- [1] involves hopscotch, contrasting with [4], which shows a garden setting with no hopscotch.\n- [1] matches the query closely, unlike [5], which shows croquet being played on a lawn.\n</contrast>\n\n<summary>\n1st=[1] (features a boy playing hopscotch as described in the query)\n2nd=[3] (another young child, but playing with a lacrosse stick instead)\n3rd=[4] (outdoor activity, unrelated to hopscotch)\n4th=[5] (features another outdoor game, croquet, unrelated to hopscotch)\n5th=[2] (features a beach soccer game, least relevant)\n</summary>\n</think>\n<answer> [1] > [3] > [4] > [5] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Nj_fvFbf56I.mp4",
"activitynet/videos/v_Vcv9P-4Tty0.mp4",
"activitynet/videos/v_CG-7jcSB5_c.mp4",
"activitynet/videos/v_DozJpP2cClI.mp4",
"activitynet/videos/v_3YdmY4oPE5s.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 guides viewers through the process of mixing, baking, and cooling a cake in her kitchen.\"\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 sequence in which a woman is in the kitchen. There is some baking, but it appears more focused on cookies. The sequence shows baking trays, cookies in an oven, and cooling steps, but there's no clear cake-making process visible. \n\n[2] This sequence features the making of an Oreo cheesecake, showing steps like mixing ingredients and assembling the cake, but no baking process, as it is likely a no-bake cake. It does not match the full query of mixing, baking, and cooling.\n\n[3] A decorated cake is being made, with activities focused on icing and adding fruits. The sequence skips the mixing and baking parts, focusing solely on decoration.\n\n[4] In this video, a woman is in a kitchen, but the activity or object involved is not clear, particularly related to cake-making. It lacks visible steps like mixing, baking, and cooling.\n\n[5] The video displays a comprehensive cake-making process which includes a woman preparing ingredients, mixing, baking in an oven, and cooling. This aligns well with the query description.\n</content>\n\n<contrast>\n- [5] clearly follows the entire cake-making process including mixing, baking, and cooling, whereas [1] focuses on cookies and lacks a step-by-step cake process.\n- [5] matches the query's baking and cooling stages, while [2] features an Oreo cheesecake, lacking baking.\n- [5] documents the entire process, contrasting with [3]'s focus on decoration only.\n- [5] presents a clear sequence of cake preparation, unlike [4] which lacks visible detail related to the query.\n</contrast>\n\n<summary>\n1st=[5] (comprehensive match with mixing, baking, and cooling a cake)\n2nd=[2] (detailed preparation but without baking a cake)\n3rd=[3] (cake decoration but no baking steps)\n4th=[1] (baking visible but focused on cookies)\n5th=[4] (unclear process and relevance)\n</summary>\n</think>\n<answer> [5] > [2] > [3] > [1] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_3OGD7if0zs8.mp4",
"activitynet/videos/v_vl7OBBTv7KY.mp4",
"activitynet/videos/v_7yndv2xOmn0.mp4",
"activitynet/videos/v_lfGmSfTjnA8.mp4",
"activitynet/videos/v_KpoH4pBp9fg.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 demonstrates how to use different objects while swimming in a pool, performing strokes and showcasing techniques as he moves back and forth.\"\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 several individuals using flotation devices in a pool. They seem to interact and perform some movements, but there is no clear demonstration of specific swimming techniques or strokes.\n\n[2] This video involves a man demonstrating exercises on a bench outside. He shows different positions and movements, but it is unrelated to swimming or pool activities.\n\n[3] The video captures a water polo game in a pool. Participants are engaged in a match rather than demonstrating individual swimming techniques.\n\n[4] The video features a single individual underwater in a pool demonstrating a technique labeled \"Backward Sink (nose holding).\" This showcases a specific swimming maneuver.\n\n[5] The video shows a man demonstrating swimming techniques with accessories, such as a pull buoy and paddles, in a pool. It highlights the use of different objects while swimming.\n</content>\n\n<contrast>\n- Item [5] beats item [4] because it explicitly shows a man demonstrating swimming techniques with different objects in a pool, which directly matches the query.\n- Item [4] beats item [1] because it clearly demonstrates a specific technique underwater, whereas item [1] involves general activities in a pool without focus on swimming techniques.\n- Item [1] beats item [3] because it involves people in a pool, albeit without clear demonstrations, while item [3] focuses on a water polo game, not on showcasing swimming techniques.\n- Item [3] beats item [2] because it is pool-related, although not a technique demonstration, whereas item [2] is entirely unrelated to swimming or pool activities.\n</contrast>\n\n<summary>\n1st=[5] (demonstrates swimming techniques with objects in a pool)\n2nd=[4] (shows a specific underwater technique)\n3rd=[1] (people in a pool but no clear technique showcase)\n4th=[3] (water polo game, not a technique demonstration)\n5th=[2] (unrelated to swimming or pool activities)\n</summary>\n</think>\n<answer> [5] > [4] > [1] > [3] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_rse-MW1helQ.mp4",
"activitynet/videos/v_FujUrbL1qc0.mp4",
"activitynet/videos/v_U0jlI70N_DQ.mp4",
"activitynet/videos/v_e51NGAPMp1s.mp4",
"activitynet/videos/v_fgBFlwM466w.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 explains and demonstrates how to properly adjust and position bike handlebars, brake levers, and shifter levers for optimal riding comfort.\"\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 demonstrates exercises on a bench, primarily related to fitness. There's no indication of bikes or bike adjustments.\n[2] A man is working on bike wheels, possibly changing a bike tire or tube, but not focusing on handlebars, brake levers, or shifter levers.\n[3] The video shows a man explaining and working on a bike, specifically focusing on adjusting the handlebars and levers, clearly aligning with the query.\n[4] The content is about arm wrestling with no connection to bike adjustments or comfort.\n[5] This video involves assembling a road bike, including setting up components, but doesn't focus on adjusting handlebars, brake levers, or shifter levers specifically for comfort.\n</content>\n\n<contrast>\n- Item [3] shows the adjustment of bike components whereas item [5] is more about assembly.\n- Item [3] focuses on handlebars and levers, which are the specific subjects of the query whereas item [2] is about changing tires.\n- Item [3] involves instructions relevant to riding comfort adjustments whereas item [1] is unrelated, focusing on fitness exercises.\n- Item [5] involves bike assembly, which is closer to the query than item [1] and item [4], but less specific than item [3].\n- Item [4] is completely unrelated, focusing on arm wrestling whereas item [2] is tangentially related due to the presence of a bike.\n</contrast>\n\n<summary>\n1st=[3] (Directly focuses on adjusting handlebars and levers for optimal riding comfort, matching the query perfectly.)\n2nd=[5] (Related to bike setup, but less specific to adjustments for comfort compared to [3].)\n3rd=[2] (Involves bike maintenance, but focuses on tires instead of handlebars and levers.)\n4th=[1] (Entirely unrelated to the query, focused on exercises.)\n5th=[4] (Completely unrelated, focused on arm wrestling.)\n</summary>\n</think>\n<answer> [3] > [5] > [2] > [1] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_FujUrbL1qc0.mp4",
"activitynet/videos/v_C_bwHYiX-Vw.mp4",
"activitynet/videos/v_nQM5LT1-ZRU.mp4",
"activitynet/videos/v_w6kBE7BHHeU.mp4",
"activitynet/videos/v_IbYlWVA8Dh8.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 engages with a group of young people playing field hockey, observes and interviews from the sidelines, receives a jersey, and then joins in the game, ultimately interacting with the players on the field.\"\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 captures a high jump event at an athletic competition. It focuses on an athlete preparing for, attempting, and clearing a high jump. There is some interaction, but it is limited to congratulatory exchanges after the jump.\n[2] This video displays a sequence where a man observes a sports game, receives a jersey, and participates in some drills on the field. It involves interaction and participation with others.\n[3] The video is about trail running, showcasing runners on various outdoor terrains, with some interview segments mixed in. There is no field hockey or related interaction as described in the query.\n[4] The video consists of various javelin throwing scenes in an athletic competition setting. There is no sideline interaction or joining in as per the query.\n[5] This video involves a futsal game taking place indoors. It contains gameplay and player interactions but no observational or sideline interviews or activities.\n\n</content>\n\n<contrast>\n- [2] is relevant as it shows a man observing, receiving a jersey, and participating in a sports activity, fulfilling the query requirements, whereas [1] is solely about a high jump event with limited interaction.\n- [1] focuses on a single sport performance in an athletic event, lacking the specific engagement and participation elements present in [2].\n- [5] involves indoor gameplay without the sideline behavior specified in the query, whereas [2] includes observing and interacting outside the game.\n- [3] and [4] depict individual sports performance and have no engagement or group interaction aspect as mentioned in the query.\n- [1], [3], [4], and [5] lack the progression from observation to active engagement present in [2].\n\n</contrast>\n\n<summary>\n1st=[2] Engages with observing, interacting, and participating in sports as specified.\n2nd=[1] Displays interaction, albeit limited, and includes sports activity.\n3rd=[5] Involves sports gameplay but lacks the observational aspect.\n4th=[3] Focuses on individual running with limited interaction.\n5th=[4] Shows athletic attempts with no relevant interaction or observation.\n\n</summary>\n</think>\n<answer> [2] > [1] > [5] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_RatbnDSuY7w.mp4",
"activitynet/videos/v_LFiGhD1OEws.mp4",
"activitynet/videos/v_lTFCUuLtUJs.mp4",
"activitynet/videos/v_FOlMVTmAeaQ.mp4",
"activitynet/videos/v_XxMzH75Bhr8.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 demonstrates how to balance and walk on a tightrope in a forest, highlighting foot placement and techniques for maintaining stability.\"\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 group of people preparing to bungee jump, visible from a platform overlooking a river. There is no tightrope or forest setting, and it does not involve tightrope walking.\n\n[2] This video depicts several people running in different outdoor settings, such as parks and dirt roads. The focus is on running and there is no demonstration of tightrope walking in a forest.\n\n[3] The video shows a man demonstrating exercises on a workout bench, likely for abdominal workouts. This does not relate to tightrope walking or forest scenes.\n\n[4] This video involves people playing croquet on a grass field. It does not have any relevance to tightrope walking or a forest environment.\n\n[5] The video illustrates a person slacklining in a forest setting. There are clear visuals of foot placement and techniques for balance on the slackline, which aligns with tightrope walking in a forest context.\n</content>\n\n<contrast>\n- [5] clearly demonstrates the action of balancing and walking on a slackline in a forest, aligning with the query, whereas [1] focuses on bungee jumping preparation.\n- [5] shows foot placement and techniques relevant to the query, whereas [2] focuses solely on running without similar context.\n- [5] is directly related to tightrope-like activity, while [3] and [4] are about exercising on a bench and playing croquet respectively, with no connection to the query.\n</contrast>\n\n<summary>\n1st=[5] (demonstrates slackline walking in a forest, closely matching the query)\n2nd=[2] (outdoor activity visible but does not relate to tightrope walking)\n3rd=[1] (outdoor setting, but focuses on bungee jumping preparation)\n4th=[3] (exercise demonstration unrelated to tightrope walking)\n5th=[4] (entirely unrelated activity: croquet on grass)\n</summary>\n</think>\n<answer> [5] > [2] > [1] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_VZ2CPdIPa30.mp4",
"activitynet/videos/v_lTFCUuLtUJs.mp4",
"activitynet/videos/v_FujUrbL1qc0.mp4",
"activitynet/videos/v_3YdmY4oPE5s.mp4",
"activitynet/videos/v_Bs3TR80BRco.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: \"Children build and play in a sandcastle by the sea, using a toy shovel and fetching water.\"\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 children at the beach constructing a sandcastle. They are using a toy shovel and bringing water to the castle, closely matching the query of building and playing in a sandcastle by the sea.\n\n[2] This video features a group of children dancing and playing in a yard, which does not relate to the query about building a sandcastle by the sea.\n\n[3] Participants are engaged in a sandcastle-building competition at the beach, with adults involved in constructing and shaping the sand. Although it involves sandcastle building, it doesn’t show children using a toy shovel or fetching water.\n\n[4] The video depicts a child playing alone in a colorful playground, unrelated to constructing a sandcastle by the sea.\n\n[5] Children are seen riding camels at a fairground, with no connection to a sandcastle activity or setting by the sea.\n</content>\n\n<contrast>\n- [1] beats [3] because it directly involves children using a toy shovel and fetching water for a sandcastle, aligning exactly with the query, whereas [3] involves adults.\n- [3] ranks higher than [2] because it involves sandcastle building at a beach, unlike the unrelated dance activity in [2].\n- [2] beats [4] as it involves multiple children playing, whereas [4] is a lone play activity unrelated to sandcastles or a beach setting.\n- [4] beats [5] as it involves children playing, albeit not on the beach, whereas [5] is a fairground camel ride not related to sand or the beach.\n</contrast>\n\n<summary>\n1st=[1] (Children building and playing in a sandcastle match all query elements)\n2nd=[3] (Sandcastle competition aligns partially with the query)\n3rd=[2] (Involves children playing but not related to sandcastles)\n4th=[4] (Child playing but not related to sandcastles or sea)\n5th=[5] (Fairground activity unrelated to the query)\n</summary>\n</think>\n<answer> [1] > [3] > [2] > [4] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_jalqZE5ekho.mp4",
"activitynet/videos/v_PjSOjB7qhcA.mp4",
"activitynet/videos/v_gYARGADLcmQ.mp4",
"activitynet/videos/v_0GpNcvAVWVg.mp4",
"activitynet/videos/v_SHfFrzy_fKM.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 play a lively game of tennis indoors, hitting the ball back and forth while moving 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] This sequence shows two men in a squash court; one holds a racket and appears to explain something or coach the other. They eventually engage in playing a game inside the glass-enclosed court.\n[2] Two men are playing an indoor racquetball game in a court. They are actively hitting the ball back and forth, showcasing movement and interaction similar to the query description.\n[3] Two individuals are playing shuffleboard on a long table in an indoor setting. This does not relate to tennis or any racquet sport.\n[4] This sequence involves children practicing basketball skills on an indoor court, focusing on lay-up techniques.\n[5] Two men play badminton indoors. They hit a shuttlecock over a net and move around a green indoor court. This involves rackets but is not tennis.\n\n</content>\n<contrast>\n- [2] shows a racquetball game with two men in an indoor court, closely matching the tennis theme in terms of action and environment, whereas [1] involves squash with more focus on instruction.\n- [1] includes explanations and squash playing but transitions from a coaching scenario, whereas [2] consistently displays a back-and-forth game dynamic.\n- [5] involves rackets but is a badminton game, not tennis, lacking the specific tennis element.\n- [3] shuffleboard and [4] basketball do not involve racket sports and are therefore less relevant to the tennis theme.\n\n</contrast>\n<summary>\n1st=[2] Two men actively play racquetball indoors, similar dynamic to tennis.\n2nd=[1] Features squash with an eventual game play, involves rackets in an indoor setting.\n3rd=[5] Involves racket play with badminton, lacks specific tennis element.\n4th=[3] Shuffleboard game, unrelated to racket sports.\n5th=[4] Indoor basketball practice, unrelated to racket sports.\n\n</summary>\n</think>\n<answer> [2] > [1] > [5] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_jzBR9ihZBbI.mp4",
"activitynet/videos/v_mhqfGS59FG0.mp4",
"activitynet/videos/v_NZ5iioSnCO8.mp4",
"activitynet/videos/v_y3Wfx-RvTpk.mp4",
"activitynet/videos/v_F7V-FqgG3T8.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 older man repairs and adjusts a bike chain outside, demonstrating the process and explaining his work as he goes.\"\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 a younger man inside a shop as he adjusts a bike, likely focusing on repairs. He seems to be working on gears but isn't seen outside or visibly explaining his work.\n\n[2] An older man is seen outdoors working closely on a bike chain. He handles the chain with tools, suggesting adjustments and repairs consistent with the query. His presence and actions strongly match the description.\n\n[3] The video shows a person welding indoors, with no connection to bicycle repairs or adjustments. The process involves heavy machinery, irrelevant to the query.\n\n[4] Displays a man working on a green bike indoors. The focus appears to be on general bike assembly or checks, but not specifically on chain adjustments.\n\n[5] This sequence features a man indoors demonstrating bicycle shifters, focusing on different bike parts than the chain. There is no outdoor setting or older man present.\n</content>\n\n<contrast>\n- [2] performs the task specified in the query outdoors, whereas [1] occurs indoors without clear chain work.\n- [3] is irrelevant due to the absence of a bicycle and focuses on welding.\n- [4] does not clearly emphasize chain work and happens indoors.\n- [5] emphasizes shifter demonstration indoors instead of chain adjustment.\n</contrast>\n\n<summary>\n1st=[2] (Match: Older man outdoors repairing bike chain clearly shown)\n2nd=[1] (Bike repair shown, but indoors and with a younger individual)\n3rd=[4] (Indoors bike work shown, not specifically chain-focused)\n4th=[5] (Focus on different bike components, no chain work)\n5th=[3] (Irrelevant welding activity)\n</summary>\n</think>\n<answer> [2] > [1] > [4] > [5] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_jbSBQL08VHA.mp4",
"activitynet/videos/v_NpKBAtQ5M6I.mp4",
"activitynet/videos/v_5Bo0gFXxDQk.mp4",
"activitynet/videos/v_5BAvlsHfTLk.mp4",
"activitynet/videos/v_nQM5LT1-ZRU.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 boys spend time together outdoors, starting with gardening and then playing a game of croquet while chatting and enjoying each other's company.\"\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 people playing a team sport on a large grassy area, which appears to be ultimate frisbee or a similar game. There is no evidence of gardening or a croquet game. \n[2] The video aligns perfectly with the query. It shows two boys starting with gardening activities, followed by setting up and playing a game of croquet. The atmosphere is relaxed, fitting the theme of spending time outdoors with company.\n[3] This video depicts a man and a child mowing a lawn in a suburban neighborhood. There's no evidence of a group of boys or croquet. \n[4] This video shows a croquet game being played, but it appears to involve older individuals rather than a group of boys. There is no sign of gardening activities.\n[5] The video depicts a toddler playing with a lacrosse stick on a wooden deck. There is no evidence of a group activity, gardening, or croquet within the video.\n</content>\n\n<contrast>\n- [2] accurately depicts both gardening and croquet with a group of boys, whereas [4] only shows croquet with older individuals.\n- [1] includes an outdoor game but not croquet or gardening, whereas [2] features both gardening and croquet.\n- [3] lacks both croquet and gardening, featuring a different outdoor activity instead.\n- [5] depicts a single child playing, which does not align with the group activity described in the query.\n- [4] involves croquet but not with a group of boys or any gardening, unlike [2], which covers both aspects.\n</contrast>\n\n<summary>\n1st=[2] (perfect match: gardening and croquet with a group of boys)\n2nd=[4] (croquet game visible but not with boys or gardening aspects)\n3rd=[1] (outdoor activity with a group, not gardening or croquet)\n4th=[3] (outdoor activity but unrelated to query specifics)\n5th=[5] (single child playing, not related to group activity)\n</summary>\n</think>\n<answer> [2] > [4] > [1] > [3] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_wUg9S2xx850.mp4",
"activitynet/videos/v_eypKc4FyZ2s.mp4",
"activitynet/videos/v_L_fvvmvV-Tc.mp4",
"activitynet/videos/v_3YdmY4oPE5s.mp4",
"activitynet/videos/v_CG-7jcSB5_c.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 gather around a table to watch and record a game as it unfolds.\"\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 bowling game in progress. Different frames display bowlers throwing balls down lanes, pins being knocked down, and scoreboards. No table or gathering around a table is evident.\n[2] This video depicts an athletics event, specifically a hammer throw. There are shots of athletes preparing, throwing the hammer, and the audience watching. No scene of men gathering around a table.\n[3] The sequence involves a weightlifting event. Athletes are lifting weights on a stage with an audience and cameras capturing the event. There isn't any gathering around a table.\n[4] This video fits the query closely. Several men are engaged around a shuffleboard table, apparently watching and possibly recording or scoring as the game progresses. The setting is casual, indicating a relevant scenario.\n[5] This video contains a hockey game with players, referees, and some crowd interactions. There is action on the ice, but no evidence of men around a table watching or recording a game.\n</content>\n\n<contrast>\n- [4] perfectly matches the query with men gathering around a shuffleboard table, whereas others feature different sports settings.\n- [1] shows a bowling scene which lacks the table-focused activity described, whereas [4] captures such a gathering.\n- [2] features an athletics event, not involving a table setting, unlike [4].\n- [3] presents weightlifting without a table scene, whereas [4] shows the relevant setup.\n- [5] revolves around ice hockey and player interaction, not around a table, unlike [4].\n</contrast>\n\n<summary>\n1st=[4] (Men gather around a shuffleboard table, aligning with the query)\n2nd=[1] (Bowling game with minimal relevance to watching around a table)\n3rd=[2] (Athletics with audience; different focus than gathering around a table)\n4th=[3] (Weightlifting event, no table-related action)\n5th=[5] (Hockey game lacking the table element)\n</summary>\n</think>\n<answer> [4] > [1] > [2] > [3] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_1t0yXLlX1XE.mp4",
"activitynet/videos/v_KnHUAc20WEU.mp4",
"activitynet/videos/v_YK5n3YTf8pI.mp4",
"activitynet/videos/v_NZ5iioSnCO8.mp4",
"activitynet/videos/v_ruHObln9mwI.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 guards march and play trumpets before a man demonstrates an elaborate shoe-polishing process involving heat and brushing, observed by soldiers.\"\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 activity on a field, possibly a tug-of-war with participants pulling on a rope. No guards or shoe-polishing process visible.\n\n[2] Depicts a pole vaulting competition at a stadium, featuring athletes preparing and jumping. No guards or shoe-polishing involved.\n\n[3] Captures a jump rope competition indoors with an audience. No guards or shoe-polishing process present.\n\n[4] Displays another group activity resembling tug-of-war with uniforms but lacks any shoe-polishing demonstration.\n\n[5] Features guards marching and playing trumpets with soldiers observing a shoe-polishing process involving heat and brushing.\n</content>\n<contrast>\n- [5] is the most relevant because it precisely matches the query with guards, trumpet playing, and a shoe-polishing demonstration observed by soldiers.\n- [1] and [4] both include group activities and uniforms but lack the shoe-polishing context.\n- [2] and [3] focus on unrelated sports activities, making them irrelevant to the query's specifics.\n</contrast>\n<summary>\n1st=[5] (exact match with marching, trumpets, and shoe-polishing)\n2nd=[1] (group activity with uniforms, no shoe-polishing)\n3rd=[4] (similar to [1], but still lacks key elements)\n4th=[2] (completely unrelated sports event)\n5th=[3] (unrelated sports event with no thematic elements from the query)\n</summary>\n</think>\n<answer> [5] > [1] > [4] > [2] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_H3cFZelOk2Y.mp4",
"activitynet/videos/v_rv9jW-t3mT4.mp4",
"activitynet/videos/v_Sma-ydx49eQ.mp4",
"activitynet/videos/v_T3ngvDyxYiE.mp4",
"activitynet/videos/v_MnzVdvbvnMc.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 gymnast prepares, performs an impressive tumbling routine at an Olympic event, receives applause and support from the crowd and her coach, and is celebrated for her performance.\"\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 weightlifting event, showing athletes preparing, lifting weights, and results on a leaderboard. No gymnastics or tumbling routines are present.\n\n[2] Displays a gymnast performing on the parallel bars at a well-attended event, showing a complete routine which culminates in the gymnast celebrating the performance with the crowd and coaches present. \n\n[3] Centers around an athletic event (discus throw), focusing on athletes, the action of throwing, and the return of results with a podium finish and celebration, but does not feature gymnastics.\n\n[4] Shows a gymnastics event with a focus on a female gymnast preparing for and performing a tumbling routine on a track, including interactions with a coach and audience. The event concludes with visible celebration.\n\n[5] Contains multiple clips of tumbling and gymnastics routines, including athlete preparations and performances at various events. Some sequences show applause and interaction with the audience, but the footage is less clear.\n</content>\n\n<contrast>\n- [4] clearly matches the query with a complete sequence of preparation, performance, and celebration at a gymnastics event, whereas [2] focuses on a parallel bar routine without evident crowd celebration.\n- [2] features a high-quality performance with visible audience but lacks the full sequence of preparation to celebration demonstrated in [4].\n- [5] contains some related content across various events but the quality and coherence are lower than [4] and [2].\n- [3] and [1] depict athletics and weightlifting events respectively, lacking relevant gymnastics content.\n</contrast>\n\n<summary>\n1st=[4] (provides a full narrative arc matching the query: preparation, performance, celebration)\n2nd=[2] (contains performing gymnast; lacks full narrative and celebration details)\n3rd=[5] (has scattered relevant gymnastics content; low clarity)\n4th=[3] (athletics content; no gymnastics)\n5th=[1] (weightlifting content; not relevant)\n</summary>\n</think>\n<answer> [4] > [2] > [5] > [3] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_uTVjevTM6V0.mp4",
"activitynet/videos/v_5xgBlI9Xx0I.mp4",
"activitynet/videos/v_BERvPz1e_AU.mp4",
"activitynet/videos/v_DEVD2oszH48.mp4",
"activitynet/videos/v_VEDRmPt_-Ms.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 engage in a fencing match on a platform while a woman observes their duel.\"\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 frames show a performance on a platform with two individuals engaging in acrobatic or tightrope actions, not fencing, and a spectator visible. It does not include fencing equipment or actions, but there is a platform presence.\n\n[2] This scene takes place in a backyard setting, featuring casual activities and social interactions among people. There's no platform, and no fencing activities are seen.\n\n[3] The indoor gym features multiple people performing dance or gymnastics routines. No fencing or platform observed.\n\n[4] Olympic gymnastics activities, focusing on pommel horse routines. It doesn't show fencing activities or a platform setting for dueling.\n\n[5] Two individuals are engaged in a fencing match on a platform. A woman is observing, fulfilling all query criteria with visible fencing equipment and actions.\n</content>\n\n<contrast>\n- [5] matches the fencing and platform setting with an observer whereas [1] involves a performance on a platform without real fencing actions.\n- [5] shows actual fencing whereas [2] is a casual outdoor scene with no relation to fencing or platforms.\n- [5] involves structured sports activity on a platform, unlike [3], which features indoor dance or gymnastics.\n- [5] directly visualizes fencing and matches the query, whereas [4] focuses on gymnastics with no relation to fencing.\n</contrast>\n\n<summary>\n1st=[5] (Direct match with fencing, platform, and observer)\n2nd=[1] (Platform setting and observer, but no fencing involved)\n3rd=[4] (Sports setting, but gymnastics instead of fencing)\n4th=[3] (Indoor activity but not related to fencing)\n5th=[2] (Unrelated outdoor social activity)\n</summary>\n</think>\n<answer> [5] > [1] > [4] > [3] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_oeFUzyNgxQU.mp4",
"activitynet/videos/v_DozJpP2cClI.mp4",
"activitynet/videos/v_NVtRGWuiEP4.mp4",
"activitynet/videos/v_Le6fNx5IWRI.mp4",
"activitynet/videos/v_J5wZTdD4XeU.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 features a group gathering outside a building and showcases a cartoon depiction of sumo wrestlers battling while a referee observes.\"\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 a person creating a Chinese painting of waterfalls and rocks. There is no depiction of sumo wrestlers or a group gathering outside.\n\n[2] This video shows an athletic event featuring a high jump. There are no sumo wrestlers or group gatherings outside depicted.\n\n[3] This video depicts an indoor activity where people are playing a game, possibly involving a floor mat resembling a sports field. There's no depiction of sumo wrestlers or an outdoor gathering.\n\n[4] The video includes animated scenes of a building with people gathering outside, followed by cartoon depictions of sumo wrestlers battling with a referee. This matches the query.\n\n[5] This video captures a live sumo wrestling match with wrestlers and a referee in a traditional setting, but lacks the cartoon depiction or outdoor group gathering described in the query.\n</content>\n\n<contrast>\n- [4] is ranked highest because it features both an outdoor gathering and a cartoon depiction of sumo wrestling, directly matching the query.\n- [5] focuses on a real sumo match and includes a referee, but lacks the cartoon style and outdoor gathering.\n- [3] shows an indoor game activity and lacks any connection to sumo wrestling or outdoor gatherings.\n- [2] is about a high jump event and is unrelated to the query's themes.\n- [1] focuses on painting, which is completely unrelated to the query.\n</contrast>\n\n<summary>\n1st=[4] (matches both aspects of the query: outdoor gathering and cartoon sumo wrestling)\n2nd=[5] (shows sumo wrestling and a referee but lacks cartoon style and outdoor aspect)\n3rd=[3] (indoor activity unrelated to the query's topics)\n4th=[2] (depicts athletic high jump, unrelated to the query)\n5th=[1] (features traditional painting, unrelated to the query)\n</summary>\n</think>\n<answer> [4] > [5] > [3] > [2] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_S47vfJ9g3Mw.mp4",
"activitynet/videos/v_VCg6_fuipp8.mp4",
"activitynet/videos/v_y0VZOcxbjRM.mp4",
"activitynet/videos/v_P9jIpcRGeOk.mp4",
"activitynet/videos/v_om1vXrpctVE.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 take turns playing guitar and interacting with the instrument.\"\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 in multiple sequences. There is no presence of a guitar or interaction with two people sharing turns.\n[2] The frames depict a violin and provide a textual guide on playing minor chords. There is no sign of guitar playing or two men interacting.\n[3] This video features two individuals preparing to play rock-paper-scissors. The focus is on their interaction, but not on guitar playing.\n[4] The video frames show individuals involved in a construction or renovation project, laying down tiles inside a room. No guitars or two men trading turns playing an instrument are present.\n[5] The series of images centers on guitar-related activity. Both a close-up of a guitar and a man operating the instrument are featured. However, the interaction does not explicitly show two men taking turns with the guitar.\n</content>\n\n<contrast>\n- Candidate [5] shows guitar-related activities, whereas [1], [2], [3], and [4] do not involve guitars.\n- [5] beats [1] because [1] involves playing a saxophone, not a guitar, whereas [5] at least shows guitar use.\n- [5] beats [2] because [2] involves a violin, not a guitar, and does not feature two people interacting with an instrument.\n- [5] beats [3] because [3] shows people preparing for rock-paper-scissors with no musical instruments involved.\n- [5] beats [4] since [4] involves construction, not a musical or guitar context.\n</contrast>\n\n<summary>\n1st=[5] (Involves guitar playing, close-up views, partial alignment with query)\n2nd=[1] (Musical setting, but wrong instrument)\n3rd=[2] (Instrument interaction, but wrong instrument)\n4th=[3] (Two-person interaction, no musical context)\n5th=[4] (Unrelated construction activity)\n</summary>\n</think>\n<answer> [5] > [1] > [2] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_CB0yYGVNcOk.mp4",
"activitynet/videos/v_MNpV9WvVPSA.mp4",
"activitynet/videos/v_AnRYS02tvRA.mp4",
"activitynet/videos/v_fVg3StD8LL0.mp4",
"activitynet/videos/v_Y-1QkIGm81w.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 shows a person riding a bull in an arena, followed by the bull charging at people and knocking them down as others watch from around the arena.\"\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 clearly shows a rodeo arena setting with a person riding a bull initially, followed by the bull running across the arena and people interacting with the scene, matching the query well.\n[2] This sequence involves equestrian activities like jumping and horseback riding, neither bulls nor events matching the query.\n[3] The frames depict an indoor sports event, specifically futsal, which is unrelated to the query scenario of a bull riding and charging event.\n[4] This video shows a rodeo scene, but it focuses on calf roping rather than bull riding or a bull charging at people.\n[5] Similar to [4], this video captures a roping event in a rodeo setting, lacking bull riding or a bull charging scene.\n\n</content>\n<contrast>\n- [1] ranks higher than [2] because it involves bull riding and subsequent charging, whereas [2] only features horse jumping.\n- [1] is above [3] because it depicts rodeo activities rather than unrelated indoor sports.\n- [1] beats [4] and [5] due to its clear depiction of bull riding and interaction, whereas [4] and [5] focus on roping.\n- [4] and [5] are ranked based on their closer relation to rodeo, yet neither includes bull charging scenes as in [1].\n- [2] is preferred over [3] due to equestrian content, which is nearer to rodeo than indoor team sports.\n</contrast>\n<summary>\n1st=[1] (depicts exact bull riding and charging)\n2nd=[4] (rodeo setting, lacks bull charging)\n3rd=[5] (rodeo setting, lacks bull charging)\n4th=[2] (equestrian, not rodeo)\n5th=[3] (unrelated indoor sports)\n</summary>\n</think>\n<answer> [1] > [4] > [5] > [2] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_F4ERtlGMFt4.mp4",
"activitynet/videos/v_2i_rotAjuoE.mp4",
"activitynet/videos/v_XxMzH75Bhr8.mp4",
"activitynet/videos/v_-X7fBYN5fBc.mp4",
"activitynet/videos/v_nnWON1EzK0o.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 surfs in the ocean, skillfully gliding over and riding waves on his surfboard.\"\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 wakesurfing behind a boat on a body of water, using a surfboard and performing tricks. This involves riding the wake produced by the boat rather than ocean waves.\n\n[2] The video captures a scene where several people are in the ocean waiting for waves to surf. Although large waves are present, no one is seen surfing them in the frames.\n\n[3] This video is unrelated to the query as it shows a person staining a fence in a backyard. There is no ocean or surfing activity present.\n\n[4] This video displays a man surfing large ocean waves skillfully. He is visible maneuvering on the surfboard, taking on challenging waves that indicate advanced surfing techniques.\n\n[5] The video depicts windsurfing activities. People are seen using windsurfing equipment on the ocean, which is a different sport from traditional surfing.\n</content>\n\n<contrast>\n- [4] is ranked higher than others because it directly shows a man surfing large ocean waves skillfully.\n- [1] involves wakesurfing which is close to the query but differs as it relies on a boat's wake rather than ocean waves.\n- [2] shows waves and surfers in the water, but lacks visible surfing action compared to [1] and [4].\n- [5] involves windsurfing, which is distinct from the query involving traditional surfing.\n- [3] is completely unrelated as it features fence staining, not any form of surfing.\n</contrast>\n\n<summary>\n1st=[4] (direct match with skillful ocean surfing displayed)\n2nd=[1] (close match with wakesurfing, showing surfing maneuvers)\n3rd=[2] (lesser match due to lack of visible surfing action, only surfers waiting)\n4th=[5] (different sport: windsurfing, not traditional surfing)\n5th=[3] (unrelated to surfing, shows fence staining)\n</summary>\n</think>\n<answer> [4] > [1] > [2] > [5] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_3svMy4Kfjd8.mp4",
"activitynet/videos/v_Cw5NDI8NnZU.mp4",
"activitynet/videos/v_FZix27qFZTc.mp4",
"activitynet/videos/v_9ZGy24Geu5k.mp4",
"activitynet/videos/v_AhWkUdag6NA.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 and a boy spend time together on a basketball court, with the man talking and the boy playing basketball.\"\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 boy is seen playing hopscotch alone in what appears to be a driveway. There is no basketball court or interaction with a man.\n[2] This sequence depicts an indoor futsal game with multiple players, unrelated to a basketball or interaction between a man and a boy.\n[3] The frames show various activities unrelated to basketball, including a computer screen, a boy indoors, and skateboarding. No basketball court or interaction is visible.\n[4] This sequence shows a man and a boy on a basketball court. The man appears to be talking, and the boy is playing basketball, matching the query.\n[5] Individuals are seen using jumping stilts in an outdoor setting, unrelated to a basketball court.\n</content>\n\n<contrast>\n- Item 4 matches the query exactly with a man talking and a boy playing basketball on a court, whereas Item 1 features a boy playing hopscotch alone.\n- Item 4 has a basketball setting which fits the query, whereas Item 2 shows a futsal environment with no basketball or family interaction.\n- Item 4 involves the right actions within a basketball context, whereas Item 3 depicts unrelated scenes like skateboarding and computer use.\n- Item 4 is directly relevant with basketball interaction compared to Item 5, which shows jumping stilts without a basketball court.\n</contrast>\n\n<summary>\n1st=[4] (matches the setting and actions)\n2nd=[1] (features a child playing, but not basketball)\n3rd=[5] (contains outdoor play, different activity)\n4th=[3] (different activities showcased, no basketball)\n5th=[2] (features futsal, wholly unrelated setting)\n</summary>\n</think>\n<answer> [4] > [1] > [5] > [3] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Nj_fvFbf56I.mp4",
"activitynet/videos/v_c6ie4KwFnAs.mp4",
"activitynet/videos/v_DXIHbnyfNGA.mp4",
"activitynet/videos/v_Kbr2nr94EB0.mp4",
"activitynet/videos/v_qxlVkDpdyOA.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 practices motocross by performing high jumps and riding his motorbike in a dusty field.\"\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 people and dogs performing agility tasks in a grassy field. This is unrelated to motocross.\n[2] The video features a high jump athletic event, focusing on an athlete jumping over a bar. No motocross or motorbikes are visible.\n[3] The video depicts motocross stunts, with a person riding a motorbike in a dusty field, performing high jumps and tricks. This matches the query well.\n[4] The video involves people outdoors, possibly a comedic scene or social event, not related to motocross.\n[5] The video includes motorbikes and riders in a race-like setting with multiple bikers on a dirt track. It somewhat matches motocross but less focus on high jumps or stunts.\n</content>\n\n<contrast>\n- [3] clearly depicts motocross activities with visible high jumps and tricks, whereas [5] focuses more on racing without distinct stunts.\n- [5] shows a motorbike-focused activity, better fitting the motocross context than [1], [2], and [4], which do not involve motorbikes.\n- [2] showcases athletic high jumps but without motorbikes, making it less relevant than any motocross content.\n- [4] involves a non-motocross outdoor social activity, less relevant than any video involving bikes.\n- [1] features a different sports activity involving dogs, irrelevant to motocross or biking.\n</contrast>\n\n<summary>\n1st=[3] (best match for motocross with stunts and jumps)\n2nd=[5] (motorbike racing activity, less on stunts)\n3rd=[2] (high jump event, no bikes but conceptually closer than others)\n4th=[4] (non-motocross social activity, unrelated)\n5th=[1] (dog agility sport, unrelated)\n</summary>\n</think>\n<answer> [3] > [5] > [2] > [4] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_e__FVkfXB-U.mp4",
"activitynet/videos/v_VCg6_fuipp8.mp4",
"activitynet/videos/v_p6jip892RdM.mp4",
"activitynet/videos/v_BJGACTBaU_Y.mp4",
"activitynet/videos/v_Fpivz0Yie_Q.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 follows a group of men engaging in different activities and hobbies in the mountains, culminating with a man speaking as the credits roll.\"\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 seems to follow a group of men engaging in different activities in a mountainous area. It concludes with a man speaking as credits roll, matching the query's description.\n\n[2] The video is primarily focused on skiing events or competitions in a snowy mountain setting. It doesn't feature a group of men engaging in various activities or any speaking parts with credits.\n\n[3] This video shows snowboarding on a mountain, focusing on individuals rather than a group of men engaging in various activities. No speaking or credits are evident.\n\n[4] The video features solo kayak paddling on a lake; it does not align with the setting or content of the query (mountains, multiple activities, speaking).\n\n[5] This video displays snowboarding in a mountain setting but lacks the element of varied activities or any speaking and credits rolling.\n</content>\n\n<contrast>\n- [1] matches the query with varied activities in the mountains and includes a speaking part with credits, whereas [2] focuses solely on skiing competitions.\n- [2] shows a skiing event, whereas [3] displays snowboarding but lacks varied activities or a speaking segment.\n- [3] involves snowboarding only, whereas [4] involves kayaking in a non-mountainous setting.\n- [4] features kayaking without the mountain setting or group activities, whereas [5] includes snowboarding but misses varied activities and a speaking part.\n</contrast>\n\n<summary>\n1st=[1] (Covers varied mountain activities and ends with speaking and credits)\n2nd=[2] (Focused on snow sports but lacks varied activities and speaking)\n3rd=[3] (Displays snowboarding in mountains, no varied activities or speaking)\n4th=[5] (Shows snowboarding, lacks varied activities and speaking)\n5th=[4] (Focused on kayaking, doesn't fit mountainous group activities)\n</summary>\n</think>\n<answer> [1] > [2] > [3] > [5] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_jHbf_E8aY1k.mp4",
"activitynet/videos/v_FNB1BMEtIvw.mp4",
"activitynet/videos/v_gBTnWp9VLz4.mp4",
"activitynet/videos/v_sQtT9sUyUwE.mp4",
"activitynet/videos/v_vu65aIIJHtU.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 play water polo in a swimming pool while others move around and interact both in and out of the water.\"\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 shows a water polo game in a swimming pool, with several players wearing caps, actively engaged in play. There are spectators and some interaction at the poolside.\n\n[2] This sequence features a diving event. An individual performs multiple dives from a platform into a pool. There is no water polo game or evident interaction between groups in or out of the pool as per the query.\n\n[3] In this sequence, many people are in a pool, but it mostly appears to be casual play or swimming without structured water polo activity. The individuals are interacting both in and out of the water, matching part of the query.\n\n[4] This sequence is a beach volleyball game with people playing on sand. There is no swimming pool or water interaction involved, which does not align with the query.\n\n[5] A group of people are in a pool playing a structured game that resembles water polo, with noticeable interaction among the players. There is evident engagement both in and around the pool area, fitting the query well.\n</content>\n\n<contrast>\n- [5] beats [1] because there is more evident interaction and movement around the pool area in addition to the water polo game.\n- [1] beats [3] because it features an actual water polo game, whereas [3] involves general pool play without structured water polo.\n- [3] beats [2] because it involves more group interaction and activity in the pool, whereas [2] focuses solely on diving.\n- [2] beats [4] because diving involves a pool setting, while [4] features beach volleyball with no water interaction.\n</contrast>\n\n<summary>\n1st=[5] (Best overall match for water polo and interaction in and around the pool)\n2nd=[1] (Clear water polo match but less interaction outside the pool)\n3rd=[3] (Good interaction but lacks structured water polo)\n4th=[2] (Diving event, involves a pool but no water polo or interaction)\n5th=[4] (Beach volleyball, no relevance to water polo or pool activities)\n</summary>\n</think>\n<answer>[5] > [1] > [3] > [2] > [4]</answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_kLflIBGCNrc.mp4",
"activitynet/videos/v_yVsOoFr61x4.mp4",
"activitynet/videos/v_xeFS0RuvVGA.mp4",
"activitynet/videos/v_SLPD1U5iDjY.mp4",
"activitynet/videos/v_7ZX3NHDmMZY.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 demonstrates waxing their leg while lying on a bed.\"\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 a person demonstrating leg waxing in a professional setting. Various shots show the waxing process, including applying wax and using strips on the leg. The person is not on a bed but appears to be on a treatment table.\n[2] This video is related to installing floor tiles. It shows a person cutting and fitting tiles, which is irrelevant to the query.\n[3] This video depicts a person performing wrestling moves on a bed. There are no scenes of waxing; instead, it involves playful interaction, unrelated to the query.\n[4] The video captures an arm-wrestling event taking place outdoors with a crowd watching. It does not involve waxing and is non-relevant to the query.\n[5] This video shows the process of waxing legs while the person is lying on a bed. It includes applying wax and using strips, closely matching the query.\n</content>\n\n<contrast>\n- [5] is the best match because it shows leg waxing on a person lying on a bed, closely aligning with the entire query.\n- [1] also features leg waxing but on a treatment table, making it slightly less relevant than [5] which fits the bed aspect.\n- [3] involves activities on a bed but no waxing, making it less relevant.\n- [4] features arm-wrestling which is entirely unrelated to the query.\n- [2] is about tile installation, which is irrelevant to waxing a leg or any bed activity.\n</contrast>\n\n<summary>\n1st=[5] (Best match: waxing on a bed, aligns with query)\n2nd=[1] (Good match: waxing demonstrated, but not on a bed)\n3rd=[3] (Some relevance: involves a bed, but no waxing)\n4th=[4] (Irrelevant: arm-wrestling event)\n5th=[2] (Irrelevant: tile installation)\n</summary>\n</think>\n<answer> [5] > [1] > [3] > [4] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_2Xfigl8xrXc.mp4",
"activitynet/videos/v_ZIBWRRBft8g.mp4",
"activitynet/videos/v_hgpekgU915k.mp4",
"activitynet/videos/v_eil9mWGJB8E.mp4",
"activitynet/videos/v_cRDXBF2RcYI.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 cheerleading team from Mill Creek High School performs an energetic and synchronized routine filled with stunts and cheers before celebrating together at 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 depicts a small group of people performing a coordinated routine in what appears to be a gymnasium. The individuals are dressed in athletic clothing and are using pom-poms, indicative of a cheerleading performance. There is movement and stunts throughout, although the setting is informal, hinting at a practice or community event rather than an official high school team performance. \n\n[2] This sequence captures an individual performing a gymnastics routine on a balance beam. The setting is a large venue with many spectators, indicating a competition. The focus is on acrobatics and balance rather than cheerleading or team dynamics.\n\n[3] The scene shows a nighttime baseball or softball game with several players on a field. The participants are actively engaged in the game, but there is no indication of cheerleading elements such as stunts or routines.\n\n[4] The sequence displays a formal cheerleading competition with a large group of cheerleaders performing a well-synchronized routine. It includes complex stunts, formations, and cheer techniques. There are banners indicating a state championship setting, making it closely aligned with the query’s mention of a cohesive and dynamic cheer performance.\n\n[5] Similar to [2], this showcases a gymnastics beam routine in a competitive environment with a large audience. The focus is on individual performance and acrobatics, not cheerleading or team interaction.\n</content>\n\n<contrast>\n- [4] is highly relevant due to depicting a full cheerleading team in a formal setting performing an advanced routine, whereas [1] features an informal and smaller performance that lacks the same complexity.\n- [1] involves cheerleading elements and synchronization but lacks the competitive and complete team environment shown in [4].\n- [2] and [5] are about gymnastics, not cheerleading, making them less relevant. [2] is similar in competitive atmosphere but lacks the cheerleader focus.\n- [3] is unrelated as it captures a sports event with no cheer performance.\n</contrast>\n\n<summary>\n1st=[4] (captures a full, formal cheerleading routine with all key elements)\n2nd=[1] (features cheerleading but in an informal setting and smaller scale)\n3rd=[2] (captures athletic performance but unrelated to cheerleading)\n4th=[5] (similar to [2], a gymnastic routine, unrelated to query)\n5th=[3] (completely unrelated as it features a baseball/softball game)\n</summary>\n</think>\n<answer> [4] > [1] > [2] > [5] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_8w9pf_Ecm_U.mp4",
"activitynet/videos/v_3j52keiQuiw.mp4",
"activitynet/videos/v_URzytLugoVw.mp4",
"activitynet/videos/v_xqYII12ta4E.mp4",
"activitynet/videos/v_ok3Kpgv0PzE.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 people of different ages take turns serving and hitting a shuttlecock while playing a game of badminton 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] This sequence shows multiple people on a badminton court, taking turns hitting a shuttlecock. The action is consistent with playing badminton, involving serving and hitting back and forth.\n[2] This sequence shows people playing croquet in a backyard. They use mallets to hit balls through hoops, which does not match the badminton query.\n[3] This video depicts people playing in a backyard, with a focus on casual outdoor activity. While there is a badminton racket visible, it lacks clarity in action.\n[4] This sequence features several people of different ages playing badminton indoors, with a clear view of serving and hitting a shuttlecock across the net.\n[5] This video shows a group of people playing croquet on a lawn. The focus is on using mallets and balls rather than badminton play.\n</content>\n\n<contrast>\n- [4] shows multiple people actively playing badminton, clearly displaying serving and hitting, whereas [1] has less diversity in action.\n- [1] focuses on badminton, displaying the movement and equipment related to a game setting, whereas [2] and [5] depict croquet.\n- [3] includes a badminton racket but lacks a clear focus on actual play compared to [1] and [4].\n- [2] and [5] are clear mismatches focused on croquet, whereas [1] and [4] include badminton gameplay.\n</contrast>\n\n<summary>\n1st=[4] (features a clear focus on multiple people of different ages playing badminton)\n2nd=[1] (showcases badminton play but with less variety in participants and action)\n3rd=[3] (includes a backyard setting and a badminton racket but lacks focused play)\n4th=[2] (depicts croquet, not matching the badminton query)\n5th=[5] (also focused on croquet, not badminton)\n</summary>\n</think>\n<answer> [4] > [1] > [3] > [2] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_9_wuEvN2OnQ.mp4",
"activitynet/videos/v_uZC3WPkta-I.mp4",
"activitynet/videos/v_DozJpP2cClI.mp4",
"activitynet/videos/v_EL47wkhiOq8.mp4",
"activitynet/videos/v_3YdmY4oPE5s.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 enjoy sliding down a snowy hill on inflated tires, with both children and adults participating in the fun.\"\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 snowy urban area with vehicles and people walking. No evidence of people sliding on inflated tires.\n[2] People are sliding down snowy tracks on inflatable tubes. Both children and adults are visible enjoying the activity.\n[3] Various winter activities occur, including people sliding on snow, though inflatable tires aren't clearly visible. There are also scenes of snowball making and playing.\n[4] The video is related to ice hockey, with scenes of players and hockey action, not related to sliding on snow.\n[5] The video appears to involve winter sports like snowboarding or skiing, but not specifically tubing on snow with inflatable tires.\n</content>\n\n<contrast>\n- [2] is the best match because it clearly shows people sliding on snow using inflatable tubes, matching the query directly.\n- [3] relates to snow activities but does not clearly show sliding on inflatable tires, making it less relevant than [2].\n- [1] is unrelated as it shows an urban setting with no sliding activity visible.\n- [5] involves winter sports like snowboarding, which doesn't match the specific activity requested.\n- [4] is entirely unrelated as it focuses on ice hockey, not snow sliding.\n\n</contrast>\n\n<summary>\n1st=[2] (clearly shows people sliding on snow with inflatable tubes)\n2nd=[3] (includes snow sliding activities but not specific to tubes)\n3rd=[5] (winter sports shown, but not the queried activity)\n4th=[1] (urban area with no relevant activities)\n5th=[4] (ice hockey, unrelated to the query)\n</summary>\n</think>\n<answer> [2] > [3] > [5] > [1] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_thvpt_lUxTI.mp4",
"activitynet/videos/v_7QvvqWJRwNo.mp4",
"activitynet/videos/v_HEuC9ZEY91E.mp4",
"activitynet/videos/v_wZ-teiWX4mg.mp4",
"activitynet/videos/v_FNB1BMEtIvw.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 performs various wakeboarding tricks, including jumping obstacles, spinning, and flipping in the air while sliding on ramps.\"\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 contains various activities, including skateboarding and kayaking, with no focus on wakeboarding tricks or obstacles.\n\n[2] This video shows a man performing wakeboarding tricks at a wake park, including jumping over obstacles and spinning in the air. It matches the query well.\n\n[3] The video features a demonstration of a wakeboarding invention, but it lacks the performance of specific tricks such as jumping, spinning, or flipping.\n\n[4] The video is unrelated, depicting a person on an exercise machine in an indoor setting, with no connection to wakeboarding.\n\n[5] This video shows wakeboarding on a lake but lacks the variety of tricks mentioned in the query, such as jumping obstacles, spinning, or flipping.\n</content>\n\n<contrast>\n- [2] shows wakeboarding tricks and park settings, matching the query because it includes jumping over obstacles and spinning, whereas [5] lacks specific trick performances.\n- [5] shows basic wakeboarding without tricks, whereas [3] includes wakeboarding with a focus on an invention.\n- [3] relates to wakeboarding but is more about a device than performing tricks, whereas [1] shows mixed sports content unrelated to the query.\n- [1] shows multiple sports but not focused on wakeboarding tricks, whereas [4] is entirely unrelated, depicting indoor exercise.\n</contrast>\n\n<summary>\n1st=[2] because it features wakeboarding tricks and obstacles.\n2nd=[5] because it shows wakeboarding but lacks tricks.\n3rd=[3] because it involves wakeboarding with focus on a device.\n4th=[1] because it shows unrelated sports activities.\n5th=[4] because it is unrelated and indoors.\n</summary>\n</think>\n<answer> [2] > [5] > [3] > [1] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_1opIvQF0m_M.mp4",
"activitynet/videos/v_hHxEBnhLsXs.mp4",
"activitynet/videos/v_wdsZn1Yu_5M.mp4",
"activitynet/videos/v_8874vqLm958.mp4",
"activitynet/videos/v_8UvKQnfXnCk.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 participate in a step workout routine at the gym and finish by raising their hands and stopping the exercise.\"\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 sequence in the frames shows people engaged in an activity in a gym, but it appears to be more of a dodgeball or group sport rather than a step workout routine. The focus is on movement with a ball, and there's no progression towards raising hands and stopping exercise in a step workout manner.\n\n[2] This sequence shows a group of people in a gym environment participating in a step workout. The presence of step platforms and synchronized movement aligns with step exercises, and it concludes with group activity, which may include a final movement such as raising hands.\n\n[3] The images depict training on monkey bars, primarily involving children and an outdoor setting. This is unrelated to a gym-based step workout routine, as it focuses on individual physical training rather than group step exercises.\n\n[4] This shows a workout instruction scenario, likely focused on abdominal exercises given the presence of crunches. While related to gym activity, it doesn't match the group dynamic or step workout focus of the query.\n\n[5] The setting involves a group of people dancing or participating in an aerobics-like exercise in a gym environment. However, there aren't visible step platforms or the raising hands ending described in the query.\n</content>\n\n<contrast>\n- [2] beats [5] because it shows visible step platforms and synchronized group movement, fitting the step workout criteria, whereas [5] suggests more of a dance routine.\n- [5] beats [1] because it involves a group exercise setting similar to a gym class, whereas [1] depicts a sport activity unrelated to the query.\n- [1] beats [4] as it involves a larger group dynamic in a gym, whereas [4] focuses on an individual exercise session.\n- [4] beats [3] because it at least takes place in a gym setting, whereas [3] is entirely out of scope with outdoor training on monkey bars.\n</contrast>\n\n<summary>\n1st=[2] (clear step workout with ending fit)\n2nd=[5] (group exercise class setting)\n3rd=[1] (group activity in a gym)\n4th=[4] (gym setting with exercise focus)\n5th=[3] (outdoor, unrelated to step workout)\n</summary>\n</think>\n<answer> [2] > [5] > [1] > [4] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_t_2jTjX3a_o.mp4",
"activitynet/videos/v_0Q9RHSEFeMI.mp4",
"activitynet/videos/v_tB_B3HVdO2I.mp4",
"activitynet/videos/v_0OxOI3sAIrM.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: \"A girl in a purple leotard performs a gymnastics routine on the balance beam, including jumps and a somersault, while an audience watches.\"\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 female athlete preparing for a track event, possibly a pole vault or long jump. She is wearing green shorts and a white top. No gymnastics or balance beam is present.\n\n[2] A male gymnast is performing on the parallel bars. He is wearing a blue leotard, and an audience is visible. This does not involve a balance beam routine.\n\n[3] A female gymnast in a purple leotard is performing on a balance beam. The sequence shows jumps and various movements, with an audience watching. This matches the query description closely.\n\n[4] Various gymnastics activities are shown, including floor and vault exercises. It is unclear if a balance beam routine with a purple leotard is included, as the frames are mixed and unclear.\n\n[5] A female gymnast performs on the balance beam in a black leotard. She executes complex moves and an audience is visible. However, she is not wearing a purple leotard.\n</content>\n\n<contrast>\n- [3] matches the query with a gymnast in a purple leotard executing a balance beam routine, whereas [5] has the right apparatus but the wrong leotard color.\n- [5] contains a balance beam routine but lacks the correct leotard color compared to [3].\n- [2] and [4] lack the apparatus focus; [2] is about parallel bars, and [4] contains mixed activities with unclear relevance.\n- [1] is unrelated as it shows a track event.\n</contrast>\n\n<summary>\n1st=[3] (Matches query with purple leotard and balance beam routine)\n2nd=[5] (Has balance beam routine, but incorrect leotard color)\n3rd=[4] (Mixed gymnastics, unclear if includes balance beam or purple leotard)\n4th=[2] (Focus on parallel bars, no balance beam)\n5th=[1] (Track event, not related to gymnastics routine)\n</summary>\n</think>\n<answer> [3] > [5] > [4] > [2] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_hS7VEMlJ4N0.mp4",
"activitynet/videos/v_ZT63nK4lexg.mp4",
"activitynet/videos/v_wlAnWPxnQ8k.mp4",
"activitynet/videos/v_VEDRmPt_-Ms.mp4",
"activitynet/videos/v_ShiBZnuxlmc.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 demonstrates and explains proper exercise techniques using a large exercise bar.\"\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 performing exercises on a mat. She is demonstrating movements like sit-ups, but there is no large exercise bar in use. This is not relevant to the query.\n[2] The video features a man demonstrating exercises on a bench. There is no large exercise bar involved, and it does not match the query.\n[3] The video shows a woman using a leaf blower in a garden setting. There is no connection to exercise techniques or a large exercise bar.\n[4] A woman is shown performing step exercises outdoors. She uses a step platform but not a large exercise bar. This is somewhat related due to the exercise theme but does not match the query for a large bar.\n[5] This video depicts a male gymnast performing on a pommel horse. It involves gymnastics skills but not the demonstration of exercise techniques using a large exercise bar, which does not fit the query.\n</content>\n\n<contrast>\n- [4] is ranked higher than [1] because it involves exercise, whereas [1] does not involve a bar.\n- [1] is ranked higher than [2] because [1] focuses strictly on exercise techniques, whereas [2] features a man and lacks coherence with the query.\n- [1] is ranked higher than [5] because [1] is exercise-focused, even though neither uses a large bar.\n- [2] is ranked higher than [5] because it involves explaining and demonstrating exercises, whereas [5] involves gymnastics.\n- [5] ranks higher than [3] as it shows gymnastic equipment and physical activity, unlike [3], which features unrelated activities.\n</contrast>\n\n<summary>\n1st=[4] (shows exercise although no bar)\n2nd=[1] (exercise shown, no bar)\n3rd=[2] (exercise-related, no bar)\n4th=[5] (athletic related, no bar)\n5th=[3] (no relation to exercise with a bar)\n</summary>\n</think>\n<answer> [4] > [1] > [2] > [5] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_zNpILhjhFbQ.mp4",
"activitynet/videos/v_FujUrbL1qc0.mp4",
"activitynet/videos/v_1P_y_DLlg-E.mp4",
"activitynet/videos/v_DxNIzbD6uD4.mp4",
"activitynet/videos/v_bmf7otW5TQo.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 bathes and grooms a small dog, carefully washing, shampooing, and handling the pet throughout the process.\"\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 bathing a dog outdoors. The sequence includes the dog standing in a tub, getting shampooed, and having water poured over it. The later frames show the dog being rinsed with a hose and finally placed in a cage.\n\n[2] The video portrays a person bathing a dog in a sink indoors. The process involves using shampoo, rinsing, and drying the dog with a towel. The event occurs on a countertop, with consistent handling of the dog throughout.\n\n[3] This video involves a person washing a large dog in a tub. The person is applying soap and scrubbing the dog, with consistent focus on washing activities. The setting is an indoor space.\n\n[4] The sequence shows a woman involved in various activities near a washing station. No clear dog bathing is visible.\n\n[5] The video appears to capture a child playing with a dog in a bath. It involves water but does not focus on grooming or careful handling of the dog.\n</content>\n\n<contrast>\n- [2] is ranked higher than [1] because it clearly shows indoors grooming with a focus on careful handling and consistent bathing.\n- [1] beats [3] since [1] has a visible sequence of washing and grooming a small dog, matching the query's focus on a small dog.\n- [3] is preferable to [5] as it involves a deliberate washing process, even though the dog is larger.\n- [5] ranks higher than [4] as it involves interaction with a dog in water, unlike [4] which lacks a specific dog bathing scene.\n</contrast>\n\n<summary>\n1st=[2] (indoor grooming, careful handling, focus on small dog)\n2nd=[1] (outdoor bathing, clear sequence on small dog washing)\n3rd=[3] (large dog washing focus)\n4th=[5] (interaction with dog in water)\n5th=[4] (no dog bathing visible)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [5] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Djlmto9iibw.mp4",
"activitynet/videos/v_y76trNI6U0Y.mp4",
"activitynet/videos/v_DXOKFXlx84M.mp4",
"activitynet/videos/v_eaR8oEBlYWk.mp4",
"activitynet/videos/v_O9qI26TB3R0.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 women demonstrate and discuss furniture polishing techniques, showcasing their results and the surrounding scenery while engaging another woman in conversation.\"\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 report and an outdoor community event with conversations but no relevance to furniture polishing.\n[2] Shows a woman alone demonstrating a product that resembles a cleaning activity, relevant to polishing but lacks the presence of two women conversing.\n[3] Depicts an interaction between two generations focused on knitting, with no relevance to furniture or polishing.\n[4] Shows two women demonstrating furniture polishing and discussing techniques. Scene includes the setting and results, matching the query.\n[5] Displays outdoor activities and social gatherings unrelated to furniture polishing.\n</content>\n\n<contrast>\n- [4] is best because it shows two women actively demonstrating and discussing furniture polishing, closely fitting the query.\n- [2] is relevant as it involves a product demonstration for cleaning/polishing, though it lacks the interaction aspect specified in the query.\n- [1] involves conversations and interactions but lacks any reference to furniture or polishing tasks.\n- [5] features unrelated outdoor activities, making it less relevant.\n- [3] focuses on a different skill (knitting) and lacks any connection to furniture or polishing.\n</contrast>\n\n<summary>\n1st=[4] (Two women demonstrate and discuss furniture polishing)\n2nd=[2] (Demonstrates related product but lacks two women interaction)\n3rd=[1] (Involves conversation but unrelated to polishing)\n4th=[5] (No connection to polishing, just social activities)\n5th=[3] (Focuses on a different activity—knitting)\n</summary>\n</think>\n<answer> [4] > [2] > [1] > [5] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_p800u2wCKbE.mp4",
"activitynet/videos/v_VQPfRCsSmQk.mp4",
"activitynet/videos/v_xxXspvK4tEY.mp4",
"activitynet/videos/v_pn1XDTlNMso.mp4",
"activitynet/videos/v_DozJpP2cClI.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 boy escapes the challenges of school and family life through drawing, dancing, and hopscotch, ultimately finding hope and connection in his creativity.\"\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 depicts a young boy at home and school, feeling isolated, but finding joy and expression through drawing and dancing. It ends with a shot of hopscotch, aligning closely with the query about escape and hope through creativity and childhood games.\n\n[2] A young boy is playing hopscotch, showing joy and engagement in the activity. However, it focuses solely on hopscotch without elements of drawing or broader creative escape.\n\n[3] A child is playing hopscotch with a parent, but the focus is primarily on the game itself, lacking the creative escape narrative and diversity of activities described in the query.\n\n[4] The video instructs on painting flowers with no depiction of a child, drawing, or dancing, and lacks narrative elements of escape or childhood.\n\n[5] A young girl is playing hopscotch in a neighborhood, similar to [2] and [3], but lacks elements related to drawing or dancing, focusing solely on the game.\n</content>\n\n<contrast>\n- [1] beats [2] because it shows the themes of drawing, dancing, and hopscotch, thus capturing the full spectrum of escape and creativity.\n- [2] beats [3] because it focuses more on the child and the hopscotch game, showing joy which aligns partly with the query about finding hope.\n- [3] beats [5] because it includes a social aspect with an adult that might hint at a connection, whereas [5] is purely focused on the activity itself.\n- [5] beats [4] because it relates to one aspect of the query, hopscotch, whereas [4] does not relate to any part of the query involving children or escape through creativity.\n</contrast>\n\n<summary>\n1st=[1] (captures full narrative of escape through drawing, dancing, and hopscotch)\n2nd=[2] (focus on hopscotch with the child's joy partially aligning with the query)\n3rd=[3] (includes hopscotch and slight social interaction but lacks full narrative)\n4th=[5] (focuses solely on hopscotch without additional creative elements)\n5th=[4] (unrelated to the themes of childhood escape through creativity)\n</summary>\n</think>\n<answer> [1] > [2] > [3] > [5] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_mBZAFJ-Q6Mw.mp4",
"activitynet/videos/v_Nj_fvFbf56I.mp4",
"activitynet/videos/v_HafSt7EDin0.mp4",
"activitynet/videos/v_fRUoSevi63M.mp4",
"activitynet/videos/v_b5GJUtjiWkQ.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 marching band gathers in the center of a crowd and performs together as onlookers watch from the sides.\"\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 depict a large marching band gathering in the center of a plaza. The band members are in uniform, holding instruments, and are surrounded by a crowd of onlookers who are standing and watching from the sides. The sequence captures their performance in a public outdoor setting, aligning well with the query.\n\n[2] This sequence shows a capoeira demonstration, where participants engage in acrobatic movements in the center of a circle formed by onlookers. While there is a crowd and a group performance, it does not match the marching band context.\n\n[3] A group of people in teal shirts and black pants are performing a choreographed routine in the center of a ring of spectators. While it shows a public performance, it lacks the specific marching band elements and instruments described in the query.\n\n[4] The video displays a musical performance with accordion players and a brass band on a stage, but it lacks the outdoor setting and crowd interaction specified in the query about a marching band.\n\n[5] People are seen dancing in the street with a crowd around them, indicative of a festive street event. However, it does not involve a marching band or the organization of such a group formation.\n</content>\n\n<contrast>\n- [1] beats [3] because it clearly features a marching band with instruments and uniforms, whereas [3] is a choreographed dance without instruments.\n- [1] beats [4] as it depicts an outdoor marching band event with crowd involvement, whereas [4] is a studio band performance.\n- [1] beats [5] because it involves a large organized group of musicians, whereas [5] focuses on street dancing.\n- [3] beats [2] and [5] due to its more organized group performance and crowd setting reminiscent of a marching band situation.\n- [2] beats [5] since it involves a structured group performance with an audience, which [5] lacks.\n</contrast>\n\n<summary>\n1st=[1] (Direct match: marching band in outdoor setting with crowd)\n2nd=[3] (Organized group performance in public setting)\n3rd=[2] (Focused group performance with an audience)\n4th=[5] (Crowd event but lacking marching band elements)\n5th=[4] (Indoor musical performance, lacks outdoor and crowd elements)\n</summary>\n</think>\n<answer> [1] > [3] > [2] > [5] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_xC5RVs9mXyM.mp4",
"activitynet/videos/v_AkeJ-lJLNkw.mp4",
"activitynet/videos/v_vX0xxYxGj-4.mp4",
"activitynet/videos/v_D5EgASje8MQ.mp4",
"activitynet/videos/v_i1llwxnI_Vo.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 boys play catch with a yellow ball in an outdoor pool, repeatedly tossing it to each other and out of the pool in a playful game.\"\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 boys standing on an outdoor court-like area, not in a pool. They are playing a game involving poles, but there is no visible yellow ball or pool-related activity.\n[2] This video features a group of boys and some adults in an outdoor pool, clearly engaging in a game involving a yellow ball. The ball is tossed around among the participants in the pool.\n[3] Displays children in the ocean and on the beach, but there is no visible yellow ball or focused game of catch.\n[4] The scene is indoors with people in a pool, but it involves floatation devices rather than a ball game. The focus is not on a yellow ball being tossed.\n[5] Shows two individuals on a beach passing a ball, but it lacks a group setting or clear indication of a playful game in a pool environment. The ball isn't discernible as yellow.\n\n</content>\n\n<contrast>\n- [2] is the best match because it shows an outdoor pool setting with a group involved in a game with a yellow ball, fulfilling all aspects of the query. \n- [5] shows a ball being tossed but lacks the group dynamic and pool setting that [2] provides.\n- [3] includes beach and ocean play but not the specific catch game with a yellow ball in a pool.\n- [4] has a pool setting but focuses on floatation devices, without a visible yellow ball game.\n- [1] features an unrelated game outside of a pool setting with no yellow ball.\n\n</contrast>\n\n<summary>\n1st=[2] (outdoor pool, group playing with a yellow ball)\n2nd=[5] (ball tossing, lacks group and pool setting)\n3rd=[3] (outdoor water setting, no focused game with a yellow ball)\n4th=[4] (pool with floatation devices, no visible yellow ball)\n5th=[1] (unrelated outdoor activity, no pool or ball)\n\n</summary>\n</think>\n<answer> [2] > [5] > [3] > [4] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_zlXvqnBqVvI.mp4",
"activitynet/videos/v_xeFS0RuvVGA.mp4",
"activitynet/videos/v_tQmsDeu1d6M.mp4",
"activitynet/videos/v_rse-MW1helQ.mp4",
"activitynet/videos/v_ymEh5T11kSQ.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 demonstrates how to create a small bag by folding, knotting, and securing a scarf with accessories placed inside.\"\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 constructing and flying a kite using sticks, string, and a garbage bag in an outdoor setting. There is no indication of creating a bag with a scarf.\n\n[2] The video depicts a man wrapping a gift with wrapping paper, yard stick, tape, and a razor blade. No scarf or bag creation is involved.\n\n[3] The video contains scenes of shot putting, with various individuals in athletic settings. This does not relate to creating a bag with a scarf.\n\n[4] This video shows a person folding and knotting a patterned scarf, placing items inside to form a small bag. This directly aligns with the query about creating a small bag with a scarf and accessories.\n\n[5] The content showcases adjusting a dog harness, demonstrating how it fits on a dog. It does not involve making a bag with a scarf.\n</content>\n\n<contrast>\n- [4] is most relevant as it directly shows folding and knotting a scarf to make a bag, matching the query, whereas others do not.\n- [1] focuses on kite construction and flying outdoors, not related to the query, whereas [4] matches perfectly.\n- [2] involves gift wrapping, missing the specific elements of scarf folding, unlike [4].\n- [3] is about shot putting, which is unrelated to the task of creating a bag.\n- [5] involves a dog harness and lacks the elements of scarf folding and bag creation.\n</contrast>\n\n<summary>\n1st=[4] (It perfectly matches the query, showing how to create a bag using a scarf.)\n2nd=[2] (Though about wrapping, it involves folding objects, somewhat close to crafting.)\n3rd=[1] (Focuses on crafting a kite, unrelated but involves creative assembly.)\n4th=[5] (Dog harness demonstration, unrelated to scarf or bag creation.)\n5th=[3] (Shot putting events, not relevant to creating a bag with a scarf.)\n</summary>\n</think>\n<answer> [4] > [2] > [1] > [5] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_AH-3mF6wWhU.mp4",
"activitynet/videos/v_c9mrRiwmhZw.mp4",
"activitynet/videos/v_Y5qDKbTkCaY.mp4",
"activitynet/videos/v_Mcutz37Bowg.mp4",
"activitynet/videos/v_oyljZJZJ7IA.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: \"Performers sing, dance, and engage with the audience through energetic choreography and dynamic stage presence during a concert performance.\"\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 dynamic concert performance with performers singing and dancing on a stage, interacting with the audience. The presence of bright lights and large audience denotes high energy and engagement, fitting the query well.\n\n[2] Displays a sequence of art creation using ink and paint to draw floral and animal shapes. This doesn't align with a concert performance, lacking any live audience or stage presence.\n\n[3] Features a dancer performing on stage, with instruments visible in the background, suggesting a live setting. However, it seems more focused on individual dance rather than an energetic concert involving audience engagement.\n\n[4] Depicts a solo dancer performing with hula hoops in various dynamic poses. While the performance is lively and engaging, it lacks elements of singing or a band setup typical for a concert.\n\n[5] Illustrates a large group dance or acrobatic performance, with performers in formation engaging in synchronized movements under stage lights. The setting is lively and dynamic, suitable for a high-energy concert context.\n\n</content>\n\n<contrast>\n- [1] beats [5] because [1] includes both dancing and singing with evident audience interaction, whereas [5] focuses on dance without clear audience engagement.\n- [5] beats [3] because [5] has a larger group dynamic, enhancing the stage presence, whereas [3] focuses on a solo performance.\n- [3] beats [4] because [3] involves more elements of stage performance with musical equipment visible, whereas [4] is a solo act with no signs of singing or a concert setup.\n- [4] beats [2] because [4] shows a performance on stage with dynamic action, aligning more with the concept of a live event, whereas [2] is about art creation.\n- [2] has no concert elements, so it ranks last.\n\n</contrast>\n\n<summary>\n1st=[1] (Engages with both singing and dancing in a large concert setting)\n2nd=[5] (Displays dynamic group performance suitable for concerts)\n3rd=[3] (Stage performance with instruments, though focused on solo dance)\n4th=[4] (Dynamic solo act but lacks concert setup)\n5th=[2] (Art creation, no concert relevance)\n</summary>\n</think>\n<answer> [1] > [5] > [3] > [4] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_cRBnLFujD3g.mp4",
"activitynet/videos/v_N_zbcVs419E.mp4",
"activitynet/videos/v_cNvTm72aGcM.mp4",
"activitynet/videos/v_NiQaDgj8Z10.mp4",
"activitynet/videos/v_HysxcMFrTkc.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 scuba divers prepare to enter the water from a boat, with one entering smoothly while the other gets stuck and struggles before making it in.\"\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 getting into inflatable tubes to float down a river. They carry the tubes to the water and then float downstream. Not related to scuba diving or entering the water from a boat.\n\n[2] Features two people in wetsuits and masks on a boat. One person smoothly enters the water with a backward roll, while the other struggles and gets stuck initially but manages to get in. Matches the query by showing the sequence of entering the water.\n\n[3] Displays a surfing scene with several surfers in the ocean waiting for and riding waves. It does not involve scuba diving or entering from a boat.\n\n[4] Shows scuba divers on a boat before entering the water and then underwater scenes, such as marine life and a diving site. While it includes preparation to enter the water, it does not clearly show one diver struggling to enter.\n\n[5] Includes a mix of underwater footage with scuba divers and marine life; however, it lacks clear footage of preparing to enter the water and showing one diver struggling.\n</content>\n\n<contrast>\n- [2] is more relevant because it clearly shows the scenario of scuba divers entering the water from a boat, with one diver struggling, matching the query exactly.\n- [4] involves scuba diving but lacks the clear struggle and smooth entry depicted in the query, whereas [2] does show this distinction.\n- [5] also involves scuba diving but does not show the entry process well, compared to [2], which clearly shows both entering scenes.\n- [3] features surfing and thus involves an unrelated activity compared to [2]’s accurate depiction of divers entering water.\n- [1] is about tubing on a river, which does not match the scuba diving setup of [2].\n</contrast>\n\n<summary>\n1st=[2] (clearly matches query with specific entry struggle and smooth entry)\n2nd=[4] (shows scuba diving and preparation, but not the specific struggle scenario)\n3rd=[5] (involves scuba diving, lacks specific entry details)\n4th=[3] (unrelated surfing scene)\n5th=[1] (unrelated tubing scene)\n</summary>\n</think>\n<answer> [2] > [4] > [5] > [3] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_6EN1q4TmPtI.mp4",
"activitynet/videos/v_25dJIdZj6Xs.mp4",
"activitynet/videos/v_Cw5NDI8NnZU.mp4",
"activitynet/videos/v_hpU50i4WBK8.mp4",
"activitynet/videos/v_PT18KK_6jmU.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 girls experience the excitement and fear of bungee jumping at an amusement park, supporting each other as they take the plunge while a pop band performs in the background.\"\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 gymnastics setting indoors where a gymnast is performing on the bars with others watching, which does not relate to bungee jumping or an amusement park setting.\n\n[2] This video captures people tubing down a river, including various scenes of people in life jackets. While adventurous, it does not match the bungee jumping context at an amusement park with a pop band.\n\n[3] The video involves children playing on a playground slide, with no connection to bungee jumping, an amusement park, or a pop band performance.\n\n[4] This video accurately portrays girls experiencing bungee jumping, with visible emotions and support from each other in an outdoor setting. It partially matches the amusement park setting, although a pop band isn't visible.\n\n[5] This video shows a group of children hitting a pinata outdoors. It does not relate to bungee jumping or an amusement park setting with a pop band.\n</content>\n\n<contrast>\n- [4] best matches the query because it actually shows girls bungee jumping and supporting each other, whereas [1], [2], [3], and [5] do not show bungee jumping or an amusement park context.\n- [2] involves adventure with water activities but lacks the specific bungee jumping theme present in [4].\n- [1], [3], and [5] are clear mismatches as they depict indoor gymnastics, playground activities, and a pinata game respectively, with no relation to the query.\n</contrast>\n\n<summary>\n1st=[4] (accurately depicts bungee jumping and the support aspect)\n2nd=[2] (involves adventure, but not bungee jumping)\n3rd=[1] (sports activity unrelated to bungee jumping)\n4th=[5] (group activity unrelated to bungee jumping)\n5th=[3] (playground activity unrelated to bungee jumping)\n</summary>\n</think>\n<answer> [4] > [2] > [1] > [5] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_mc5XVTo5PVk.mp4",
"activitynet/videos/v_OYPyiUC4h68.mp4",
"activitynet/videos/v_C4V6fqELvPY.mp4",
"activitynet/videos/v_Hev5Vj0Jjuk.mp4",
"activitynet/videos/v_WJ4tCIdAgVo.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 navigates a rough river in his kayak, faces various challenges including overturning and a waterfall, and ultimately makes it safely to shore.\"\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 group of people whitewater rafting, not kayaking. They navigate a rough river, but there’s no focus on an individual facing a waterfall or overturning. \n\n[2] This video involves rafting with multiple people. It lacks the individual kayaking focus and specific challenges like overturning or a waterfall.\n\n[3] The video follows an individual kayaking through rough waters, encountering a waterfall, and overturning. It ends with the man safely on calm waters, closely aligning with the entire query.\n\n[4] The video shows a lone kayaker navigating turbulent waters. There’s visible challenge with rough rapids and the kayaker eventually navigates through the scenario.\n\n[5] This video depicts a kayaker in challenging whitewater conditions, likely facing rapids, but it doesn’t clearly show overturning or a waterfall. It ends with the kayaker still in the river.\n</content>\n\n<contrast>\n- [3] beats [4] because it clearly showcases overturning and a waterfall scene, while [4] lacks these specifics.\n- [4] beats [5] since [4] maintains focus on a single kayaker facing challenges, whereas [5] lacks clear progression and outcome.\n- [5] beats [1] because it features kayaking and challenging waters, unlike [1] which shows rafting.\n- [1] beats [2] as both are about rafting, but [1] has more visible whitewater action which is closer to the theme of navigating challenges.\n</contrast>\n\n<summary>\n1st=[3] (Detailed sequence with kayak, waterfall, overturn, and successful navigation)\n2nd=[4] (Kayak and rough rapids with individual focus, but less specific events)\n3rd=[5] (Shows kayaking challenges but lacks overturn/waterfall)\n4th=[1] (Rafting video with some challenges but no kayak or overturning)\n5th=[2] (Rafting with minimal relation to individual challenges of the query)\n</summary>\n</think>\n<answer> [3] > [4] > [5] > [1] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Z0oEaM7h6jM.mp4",
"activitynet/videos/v_I9_eKiGZPxc.mp4",
"activitynet/videos/v_A1SuSeqDTEI.mp4",
"activitynet/videos/v_L1XpfS1RCzE.mp4",
"activitynet/videos/v__3lgJ6yp7o8.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 fitness class follows an instructor through coordinated exercises, culminating with on-screen text.\"\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 fitness class with participants following an instructor. The exercises involve step platforms, and the sequence culminates with on-screen text. The presence of \"NOW FIT\" further indicates a structured class setting.\n\n[2] Displays a group fitness class with participants wearing athletic attire, following an instructor. The sequence suggests a step workout, and the participants are synchronized, indicating a coordinated exercise session. \n\n[3] Consists of a one-on-one fitness instruction session focusing on performing the perfect crunch. There is no group exercise component, and it lacks the thematic focus of a coordinated class.\n\n[4] Primarily showcases features of fitness equipment with no group class or exercise instruction. It does not match the query focused on coordinated group exercises.\n\n[5] Depicts a spin class with participants exercising on stationary bikes. It suggests a group fitness activity, but it lacks the step exercise and coordination specifics emphasized in the query.\n</content>\n\n<contrast>\n- [1] beats [2] because it shows a fitness class with step exercises and culminates with on-screen text.\n- [2] beats [5] because it shows step exercises and synchronized movements, whereas [5] involves spin cycling without step platforms.\n- [5] beats [3] because it involves a group exercise setting, whereas [3] only shows individual instruction.\n- [3] beats [4] because it involves a fitness routine, whereas [4] only shows equipment features.\n</contrast>\n\n<summary>\n1st=[1] (Shows a coordinated fitness class with step exercises culminating in on-screen text)\n2nd=[2] (Shows a coordinated group fitness class with step exercises)\n3rd=[5] (Depicts a group spin class)\n4th=[3] (Focuses on individual exercise instruction)\n5th=[4] (Showcases fitness equipment features)\n</summary>\n</think>\n<answer> [1] > [2] > [5] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_IsVpeZoK5FI.mp4",
"activitynet/videos/v_0Q9RHSEFeMI.mp4",
"activitynet/videos/v_0OxOI3sAIrM.mp4",
"activitynet/videos/v_JviFZ5Cliug.mp4",
"activitynet/videos/v_Q5kU3DUXXZg.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 basketball player practices shooting hoops under a net, with text appearing on the screen as the video concludes.\"\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 athletics event, possibly javelin or shot put, with no basketball content. Audience and judges are visible, but not relevant to the query.\n[2] Features various athletes in a track and field setting, not related to basketball. The setting includes multiple athletes and spectators.\n[3] Depicts a child playing hopscotch. A basketball is visible but not actively used, making it unrelated to practicing basketball shots.\n[4] Shows an indoor sports game, possibly futsal or indoor soccer, featuring players wearing jerseys. It's not basketball and lacks shooting hoop action.\n[5] Clearly depicts a basketball player practicing lay-up shots with visible hoops and structured drills. Text appears, indicating instruction on \"The Mikan Drill,\" matching the query about practice and text concluding the video.\n</content>\n\n<contrast>\n- [5] matches the query with a player practicing basketball and text onscreen, whereas [1], [2], [3], and [4] do not feature basketball at all.\n- [3] has a basketball present but lacks active shooting or practice, whereas [5] shows active shooting drills.\n- [4] involves a different sport, failing to match the domain of the query, unlike [5].\n- [1] and [2] are focused on athletics events, lacking any basketball context or practice action, unlike [5].\n</contrast>\n\n<summary>\n1st=[5] (Clearly matches with basketball practice, hoops visible, and text instruction)\n2nd=[3] (Basketball visible but not actively used, irrelevant to hoops practice)\n3rd=[4] (Different sport altogether, not basketball)\n4th=[1] (Athletics event, no relation to basketball)\n5th=[2] (Another athletics event, no connection to basketball)\n</summary>\n</think>\n<answer> [5] > [3] > [4] > [1] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_GU_8hjy79VQ.mp4",
"activitynet/videos/v_BERvPz1e_AU.mp4",
"activitynet/videos/v_HafSt7EDin0.mp4",
"activitynet/videos/v__icW_MAe7Po.mp4",
"activitynet/videos/v_DmTI8y6Anvo.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 thoroughly cleans a room by vacuuming various surfaces, removing dirt and hair, and emptying the vacuum afterward.\"\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 demonstrating the use of a vacuum cleaner, showing various tools, and offering advice, but doesn't include a woman cleaning a room thoroughly as described in the query.\n\n[2] This sequence shows a woman vacuuming a room, focusing on vacuuming various surfaces, inspecting the vacuum, and emptying it, which matches closely with the query.\n\n[3] The sequence depicts a man working on a wall, likely involved in construction or renovation, which is unrelated to the query about vacuum cleaning.\n\n[4] The sequence focuses on fitness equipment and a trainer, unrelated to vacuum cleaning or room cleaning.\n\n[5] This sequence shows two individuals sanding or applying material to walls, likely in a construction setting, and is unrelated to vacuum cleaning.\n</content>\n\n<contrast>\n- [2] better matches the query because it shows a woman vacuuming and emptying the vacuum, whereas [1] shows a man demonstrating vacuum cleaners.\n- [2] involves vacuuming, whereas [3] and [5] involve construction work, which is unrelated to the query.\n- [2] focuses on cleaning activity, which [4] does not, as it focuses on fitness equipment.\n</contrast>\n\n<summary>\n1st=[2] (Thorough vacuuming and emptying by a woman; matches query)\n2nd=[1] (Features vacuum cleaning demonstration but not by a woman)\n3rd=[3] (Shows wall work unrelated to cleaning)\n4th=[5] (Depicts wall sanding unrelated to vacuuming)\n5th=[4] (Focus on fitness equipment, unrelated to cleaning)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [5] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_C1TWr5XH8b0.mp4",
"activitynet/videos/v_tr8JRPrfAN0.mp4",
"activitynet/videos/v_2KYGAq5qQCI.mp4",
"activitynet/videos/v_jC1He93cebg.mp4",
"activitynet/videos/v_TrtuuE1K8pk.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 heated hockey game escalates into a chaotic brawl involving players, coaches, and spectators, before ending with the red team celebrating 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] Displays an indoor soccer/football match with players in yellow and red jerseys. There are scenes of gameplay, penalties, celebrations, and officials with trophies. The focus is on a match rather than a brawl or hockey game.\n[2] Contains various scenes of ice hockey gameplay, mainly with players in blue and white jerseys. It showcases typical hockey action but lacks the elements of a chaotic brawl, coaches, spectators, or team celebration.\n[3] Shows elements closest to the query: a hockey game with visible on-ice confrontations involving players and what seems like coaches or officials. There is a progression showing players in red celebrating together at the end.\n[4] Features a curling event with predominantly red-attired athletes and spectators. It captures gameplay and spectator reactions but lacks any element of a hockey game or a brawl.\n[5] Consists of ice hockey scenes focused on player injuries and hits. There is no content showing a brawl, the involvement of coaches or spectators, or a celebratory ending by a red team.\n\n</content>\n\n<contrast>\n- [3] captures a hockey brawl with on-ice chaos and ends with a red team's celebration, matching the query closely. \n- [2] shows hockey gameplay but lacks the chaos and celebration described in the query.\n- [5] focuses on injuries without depicting the query's described brawl or team celebration elements.\n- [1] is centered around an indoor soccer match, not relevant to the hockey query.\n- [4] displays curling, which is unrelated to the query's hockey focus.\n</contrast>\n\n<summary>\n1st=[3] (captures the chaotic brawl and red team celebration)\n2nd=[2] (shows hockey action but lacks brawl and celebration)\n3rd=[5] (shows hockey incidents but not a brawl or celebration)\n4th=[1] (displays soccer, not relevant to hockey)\n5th=[4] (shows curling, unrelated to query)\n</summary>\n</think>\n<answer> [3] > [2] > [5] > [1] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_ynda_PQe_CM.mp4",
"activitynet/videos/v_rF1AqlLjECk.mp4",
"activitynet/videos/v_ruHObln9mwI.mp4",
"activitynet/videos/v_31KEa5VhvPs.mp4",
"activitynet/videos/v_CEEJ1gzdmAk.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 athletes compete in a canoe race along a marked waterway, finishing as they cross the final buoys.\"\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 an individual on stilts in a grassy park area, not related to canoe racing.\n\n[2] The video depicts a person in a canoe paddling through water, but only one participant is visible throughout, with no clear indication of a race or marked waterway.\n\n[3] This video features windsurfing, with individuals on boards on the ocean. It does not depict canoeing or a race along a marked waterway.\n\n[4] This video shows multiple participants in canoes paddling along a marked waterway with buoys visible, indicating a race setting and crossing a finish line marked by buoys.\n\n[5] The video includes several canoeists paddling on a river in snowy conditions. It shows a more casual setting rather than a formal race.\n</content>\n\n<contrast>\n- [4] is the best match because it clearly depicts a canoe race along a marked waterway with multiple participants and buoys indicating a finish line, fulfilling the query requirements.\n- [5] is ranked higher than [2] because it shows multiple canoeists, though not a formal race, while [2] only shows a single paddler.\n- [2] is better than [3] because it at least involves a canoe, whereas [3] shows windsurfing, not canoeing.\n- [3] is better than [1] because it involves water sports, while [1] does not involve water at all.\n</contrast>\n\n<summary>\n1st=[4] (clearly depicts a group canoe race along a marked waterway, finishing at buoys)\n2nd=[5] (shows multiple canoeists paddling, though not a formal race)\n3rd=[2] (depicts paddling but with a single participant, not a race)\n4th=[3] (features windsurfing, not canoeing)\n5th=[1] (unrelated to canoe racing, features a person on stilts)\n</summary>\n</think>\n<answer> [4] > [5] > [2] > [3] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_NLTd7qCLT7o.mp4",
"activitynet/videos/v_sQtT9sUyUwE.mp4",
"activitynet/videos/v_AhWkUdag6NA.mp4",
"activitynet/videos/v_8dXbbJWFEJo.mp4",
"activitynet/videos/v_sgwg1C3i-Vw.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: \"Teams of boys in different colored uniforms compete in indoor soccer matches watched by spectators, with goal attempts and successful plays throughout the game.\"\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 an indoor futsal match with teams of boys in blue and red uniforms. There are several goal attempts and plays throughout, indicating active gameplay. Spectators can be seen watching from the sidelines.\n\n[2] The images depict a beach soccer match, not indoor soccer. Players are on a sand field with different segments showing refereeing and various plays. The setting is distinctly different from indoor soccer.\n\n[3] This video features an indoor futsal match with teams in yellow and red uniforms. It shows goal attempts, successful plays, and players actively engaging in the match with visible spectators.\n\n[4] Various elements do not match the query well, with segments showing news anchors and interviews rather than focusing solely on the match. Portions of the video include indoor soccer, but it is interspersed with unrelated content.\n\n[5] This sequence depicts a dodgeball game in a gym setting with adult participants, not an indoor soccer match with boys. The activities and uniforms do not match the query.\n</content>\n\n<contrast>\n- [1] ranks higher than [3] because it focuses more specifically on a futsal match involving boys, aligning more closely with the query.\n- [3] ranks higher than [4] because it has consistent indoor soccer content without interruptions from unrelated footage.\n- [4] ranks higher than [5] because it contains some relevant indoor soccer footage, whereas [5] is entirely about dodgeball.\n- [5] is ranked lower because it shows a dodgeball game, not soccer, and does not match the query.\n- [2]'s beach soccer setting is not relevant to the indoor context required by the query, so it is ranked lower.\n</contrast>\n\n<summary>\n1st=[1] (direct match with query featuring boys in indoor soccer with spectators)\n2nd=[3] (high relevance with futsal match, but less focused on boys)\n3rd=[4] (some indoor soccer content but includes unrelated footage)\n4th=[5] (entirely dodgeball, not relevant to the query)\n5th=[2] (beach soccer does not match the indoor requirement)\n</summary>\n</think>\n<answer> [1] > [3] > [4] > [5] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_c6ie4KwFnAs.mp4",
"activitynet/videos/v_dRiIYsAmNAc.mp4",
"activitynet/videos/v_XxMzH75Bhr8.mp4",
"activitynet/videos/v_Mo9-1cuMVtQ.mp4",
"activitynet/videos/v_t_2jTjX3a_o.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: \"Four bikers compete in a race, navigating jumps and curbs on the track, with one rider falling at the finish line before getting up and walking 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] This grid shows a track and field event focusing on men's triple jump. There are several frames of athletes preparing, running, and completing their jumps. No bikes or racing elements are present.\n\n[2] This grid captures a BMX race with multiple riders on a track that includes jumps and curves. The riders navigate through a series of obstacles, which matches the concept of a bike race. However, the fall and recovery at the finish line is not depicted clearly.\n\n[3] This grid centers around a javelin throwing event. Athletes are shown in various stages of preparing and throwing the javelin. There are no bikes or races involved.\n\n[4] This grid features a casual or humorous scene of people in a grassy field, possibly in a schoolyard. No competitive biking or race-related activities are visible.\n\n[5] This grid portrays a bike racing event. It shows multiple riders competing on a track with varying terrain, including jumps. There is a fall at the finish line, followed by the individual getting up, which closely matches the query.\n</content>\n\n<contrast>\n- [5] beats [2] because it captures the fall at the finish line and the recovery, closely reflecting the query, whereas [2] lacks the fall-and-recovery moment.\n- [2] beats [1] because it involves a BMX race with jumps and an outdoor track atmosphere, whereas [1] shows an unrelated triple jump event.\n- [1] beats [3] because the triple jump is closer to competitive sports than the javelin throw in terms of dynamic action, though both are off-topic.\n- [3] beats [4] because it is at least a sporting event (javelin), whereas [4] is unrelated and informal.\n- [4] is ranked lowest as it does not contain any elements related to competitive biking or a race.\n</contrast>\n\n<summary>\n1st=[5] (Captures a bike race with a fall and recovery at the finish line.)\n2nd=[2] (Depicts a BMX race with jumps but lacks the finish line fall and recovery.)\n3rd=[1] (Shows a triple jump event unrelated to biking.)\n4th=[3] (Displays a javelin event, further-off-topic but more athletic than [4].)\n5th=[4] (Completely unrelated casual scene, no race or bikers.)\n</summary>\n</think>\n<answer> [5] > [2] > [1] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_cwQQj3VZLC8.mp4",
"activitynet/videos/v_vWULU8NttX8.mp4",
"activitynet/videos/v_sO7ZGxFwbm8.mp4",
"activitynet/videos/v_BJGACTBaU_Y.mp4",
"activitynet/videos/v_OgaurUPiWYI.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 children enjoy playing and washing in a tub, interacting with toys and a helpful adult while smiling 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] Two children are in a bathtub surrounded by floating toys. The children appear joyful, playing and interacting with each other and looking at the camera. The environment suggests a playful and comfortable atmosphere, with visible water and colorful toys.\n\n[2] A child is indoors, sitting on a carpet with various household objects. There is no bathtub or water visible. The child is interacting with the items, but no adult or signs of bathing are present.\n\n[3] Two children are washing dishes in a kitchen sink with an adult aiding them. The scene has soapy water, and the children seem engaged and having fun, although this is a kitchen setting, not a bathing one.\n\n[4] A dog is being washed in an outdoor tub by an adult. There's clear interaction and washing, but it involves a dog, not children. The scenario lacks the playful child activity described.\n\n[5] A group of children is playing outdoors at what appears to be a party with some adult supervision. There is no tub or washing activity involved, as it is more of a general outdoor gathering.\n</content>\n\n<contrast>\n- [1] matches the query because it involves two children in a bathtub, playing with toys and smiling, matching the described activity of enjoying washing with adult interaction.\n- [3] involves children with water and supervision but in a kitchen setting, whereas [1] occurs in a bathtub, which is more relevant to the query.\n- [2] only involves one child and no water/tub scenario, whereas [1] has multiple children in a tub.\n- [4] involves a bath that applies to a dog, not children, whereas [1] involves the correct subject: children.\n- [5] involves play but in an outdoor setting without focus on washing or a tub, making [1] more relevant.\n</contrast>\n\n<summary>\n1st=[1] (Two children in a bathtub, smiling and playing with toys, matches query perfectly)\n2nd=[3] (Children interacting with water and an adult but in a kitchen, not a bath)\n3rd=[2] (Single child playing indoors with no water or tub)\n4th=[5] (Children playing outdoors without water involvement)\n5th=[4] (Focus is on washing a dog, not involving children bathing)\n</summary>\n</think>\n<answer> [1] > [3] > [2] > [5] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_AWAMhmc08Cw.mp4",
"activitynet/videos/v_hjDGz9OhwvM.mp4",
"activitynet/videos/v_XbSQ_7vh7yc.mp4",
"activitynet/videos/v_DXOKFXlx84M.mp4",
"activitynet/videos/v_3fg_CB3rBSk.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 gather to play darts, cheer each other on, and enjoy drinks 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 foosball game with people playing and interacting with the table. There are no visible darts or drinks.\n[2] This video features a young child playing darts alone. The setting is a room with a dartboard.\n[3] This is a professional darts match with two competing players and a crowd. There's no visible social gathering or drinks, but it fits the theme of darts.\n[4] People are playing a field sport outside. There are no visible darts, cheers, or drinks.\n[5] People are gathered outdoors, playing a casual game and sitting together, possibly drinking. There's social interaction but no darts.\n</content>\n\n<contrast>\n- [3] beats [2] because it shows people playing darts and has a cheering crowd, aligning with the social aspect.\n- [2] beats [5] because it directly involves playing darts, whereas [5] lacks this component.\n- [5] beats [1] because it involves social gathering and possible drinks, fitting more closely with the query.\n- [1] beats [4] because it involves a game setting, albeit not darts, whereas [4] is unrelated.\n- [4] is least relevant as it shows outdoor sports with no connection to darts or social gathering with drinks.\n</contrast>\n\n<summary>\n1st=[3] (professional darts scene with players and crowd)\n2nd=[2] (solo darts playing, no social gathering)\n3rd=[5] (outdoor gathering and socializing, lacks darts)\n4th=[1] (indoor game setting but no darts)\n5th=[4] (outdoor sport unrelated to darts or socializing)\n</summary>\n</think>\n<answer> [3] > [2] > [5] > [1] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_HJZj6z1dyT4.mp4",
"activitynet/videos/v_JKVOi9DE8Hs.mp4",
"activitynet/videos/v_KRz3aBw_TlQ.mp4",
"activitynet/videos/v_wUg9S2xx850.mp4",
"activitynet/videos/v_DozJpP2cClI.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 performs a belly dance for an audience, including a child and another woman, who watch and share a hug as the dance unfolds.\"\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 sumo wrestling match with no indication of a belly dance or audience interaction as described in the query.\n\n[2] Contains scenes of an indoor soccer game with players and a match in progress, lacking any dance performance or relevant audience.\n\n[3] Displays a woman performing a dance in a room with festive lights. In some frames, a seated person is visible, potentially an audience, but there's no clear interaction or presence of a child as specified in the query.\n\n[4] Captures an outdoor activity with multiple children and adults, possibly a game, but there's no indication of a dance performance or relevant audience interaction mentioned in the query.\n\n[5] Features a dance performance in a gym or auditorium, with an audience visible. However, it does not focus specifically on a belly dance or detailed audience interaction including a hug.\n</content>\n\n<contrast>\n- [3] is primarily focused on a solo dance performance with some audience presence, making it more relevant than others.\n- [5] depicts a dance performance in front of an audience but lacks the specific elements of belly dancing and audience interaction seen in [3].\n- [1], [2], and [4] do not contain dance performances or relevant actions.\n</contrast>\n\n<summary>\n1st=[3] (Dance performance closely matches the query; potential audience visible)\n2nd=[5] (Dance in front of audience; lacks specificity to belly dance and interaction)\n3rd=[4] (Outdoor activity, unrelated to dance)\n4th=[1] (Sumo wrestling, no dance or relevant audience interaction)\n5th=[2] (Indoor soccer game, no dance or audience fit)\n</summary>\n</think>\n<answer> [3] > [5] > [4] > [1] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_slUvIngXl4g.mp4",
"activitynet/videos/v_XxMzH75Bhr8.mp4",
"activitynet/videos/v_afORRx0QZgI.mp4",
"activitynet/videos/v_WJ4tCIdAgVo.mp4",
"activitynet/videos/v_8w9pf_Ecm_U.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 sits in front of the camera, repeatedly sipping tea and ultimately making a menacing expression.\"\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 appears seated indoors, with curlers in her hair, interacting with a desktop computer. The sequence progresses with another person joining her, concluding with a light-hearted expression. No tea-drinking or menacing expression is observed.\n\n[2] A woman is captured in close-up shots, repeatedly sipping tea from a cup, with expressions progressing to what may be interpreted as menacing. The sequence clearly shows the transition from sipping to a wider, intense look.\n\n[3] The video depicts a woman engaged in gift-wrapping activities, with elaborate wrapping setups and decorations. No tea-drinking or menacing expression is involved.\n\n[4] A woman is engaged in peeling fruits indoors. The sequence shows her focus on the task, with no relation to sipping tea or any menacing expressions.\n\n[5] A woman is presenting products while seated on a sofa, holding cleaning items. The sequence revolves around demonstrating the product, without any tea-drinking or significant change of expression to menacing.\n</content>\n\n<contrast>\n- Video [2] matches the query as it shows a woman sipping tea and making a menacing expression whereas others do not.\n- Video [1] shows indoor activity but lacks tea-sipping or menacing expression, whereas [2] does both.\n- Video [3] includes gift wrapping but no tea-drinking or menacing expressions, unlike [2].\n- Video [4] involves peeling fruits without any tea-drinking or menacing expressions, making [2] more relevant.\n- Video [5] is a product demonstration without the specific actions needed, whereas [2] fits the query actions.\n</contrast>\n\n<summary>\n1st=[2] (Shows a woman sipping tea and making a menacing expression)\n2nd=[1] (Indoor setting with a woman, lacks tea-drinking/menacing expression)\n3rd=[5] (Shows a woman, lacks tea-drinking/menacing expression)\n4th=[3] (Unrelated to tea-drinking; focused on gift-wrapping)\n5th=[4] (Focused on peeling fruits, unrelated to tea-drinking)\n</summary>\n</think>\n<answer> [2] > [1] > [5] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_NS8qkB39x9U.mp4",
"activitynet/videos/v_UdBFm97tOJs.mp4",
"activitynet/videos/v_n5xfkpQ8PWM.mp4",
"activitynet/videos/v_UzCeMCr0lM8.mp4",
"activitynet/videos/v_VQPfRCsSmQk.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 relax and float down a scenic river in tubes, enjoying the peaceful natural surroundings.\"\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 people floating down a river in tubes, surrounded by lush greenery. It captures a relaxed atmosphere as the participants enjoy a peaceful ride, meeting the query requirements of people floating down a scenic river in tubes.\n\n[2] The scene involves individuals skating on grass with stilts near a pond and fountain. It does not include any river tubing activities and does not fit the query.\n\n[3] This scene includes brief moments with people in water activities, but mainly emphasizes public speaking and promotional content. It lacks the relaxed, scenic view of tubing on a river as per the query.\n\n[4] The frames focus on individuals tubing down river rapids, with a sense of adventure rather than relaxation. The rapid nature slightly diverges from the peaceful theme of the query.\n\n[5] Depicts people canoeing on a calm river, surrounded by nature. While it maintains the peaceful natural surroundings, the focus on canoeing instead of tubing does not fully match the query.\n</content>\n\n<contrast>\n- [1] best matches the query through clear scenes of relaxed river tubing, whereas [4] has a more adventurous tone with rafting.\n- [4] matches more closely than [5] as the activity involves tubes, although more adventurously. \n- [5] outerclasses [3] through its calm river setting, although focused on canoeing.\n- [3] includes some river activity but is diluted with unrelated content, whereas [2] is unrelated entirely.\n</contrast>\n\n<summary>\n1st=[1] (Clear depiction of relaxed tubing in a scenic river setting)\n2nd=[4] (Adventure tubing fits query less but still beats non-tubing entries)\n3rd=[5] (Peaceful river setting matches tone, but activity differs)\n4th=[3] (Some relevant scenes, but overall context diverges)\n5th=[2] (Unrelated activity, no river tubing context)\n</summary>\n</think>\n<answer> [1] > [4] > [5] > [3] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Uk_VxavVjP0.mp4",
"activitynet/videos/v_NLTd7qCLT7o.mp4",
"activitynet/videos/v_I9_eKiGZPxc.mp4",
"activitynet/videos/v_b_v5uGSheEw.mp4",
"activitynet/videos/v_tl3wQEusmj8.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 work together painting a fence by a lakeside as the camera captures their activity and the surrounding scenery.\"\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 primarily shows people engaging in various activities such as hitting pinatas and participating in outdoor games. No painting or lakeside setting is visible. \n[2] The video depicts a man setting up and working around camping tents. There is no visible painting or a lakeside environment.\n[3] Two men are shown playing croquet in a garden. There’s no indication of painting activity or a lakeside setting.\n[4] The video captures scenes of a white fence near a lakeside. Two men are seen working on the fence, suggesting they might be painting it, which aligns with the query requirements.\n[5] Individuals are seen playing badminton and gathering in a backyard. There is no fence painting or lakeside scenery visible.\n</content>\n\n<contrast>\n- [4] is relevant because it shows two men interacting with a fence near a lakeside, likely painting it, whereas [3] involves playing croquet with no painting or lakeside context.\n- [4] shows actual fence-related activity near a lakeside, whereas [2] captures camping activities with no relevant scene or action.\n- [4] includes visible painting activity which is absent in [1] and [5], both focusing on unrelated leisure activities without lakeside views.\n</contrast>\n\n<summary>\n1st=[4] (clearly shows fence work by a lakeside)\n2nd=[2] (outdoor setting but lacks lakeside or painting activity)\n3rd=[5] (outdoor scene, but no painting or lakeside context)\n4th=[3] (involves two men but focuses on croquet, not painting)\n5th=[1] (displays pinata and games, unrelated to the query)\n</summary>\n</think>\n<answer> [4] > [2] > [5] > [3] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_ZqkJAD0CGBU.mp4",
"activitynet/videos/v_rjgu0ucCW7U.mp4",
"activitynet/videos/v_eypKc4FyZ2s.mp4",
"activitynet/videos/v_pK3OSevBqDw.mp4",
"activitynet/videos/v_DozJpP2cClI.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 white tank top plays the bagpipes for a man in a pink shirt in a parking lot, and he applauds and hugs her after her performance.\"\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 musicians, likely performing in public, but it does not show a bagpipe, a woman in a white tank top, or relevant interactions. The setting, with sightseeing buses, suggests a tourist area rather than a parking lot.\n\n[2] This video is centered around a car wash scene featuring people in swimwear, which is unrelated to music performances or the suggested query setting.\n\n[3] The video involves an arm-wrestling competition at a table, with spectators watching, and does not include any elements related to music or bagpipe performances.\n\n[4] A woman in a white tank top plays the bagpipes for a man in a pink shirt in a parking lot. The sequence shows her playing, the man applauding, and embracing her, matching the query exactly.\n\n[5] This video shows a man interacting with a car in a snowy parking lot, performing maintenance or inspection. The scene lacks music elements completely.\n</content>\n\n<contrast>\n- [4] is a precise match as it depicts a woman in a white tank top playing the bagpipes in a parking lot for a man in a pink shirt, who later applauds and hugs her, whereas [1], [2], [3], and [5] do not show bagpipes or the specified interaction.\n- [1] involves musicians but doesn't align with the detailed setting and interaction of the query, whereas [4] does.\n- [2] and [5] have no musical performance or interaction relevant to bagpipes, reducing their relevance compared to [4].\n- [3] focuses on competitive arm-wrestling, lacking any connection to music or bagpipe performances.\n</contrast>\n\n<summary>\n1st=[4] (Exact match with the query details)\n2nd=[1] (Musical elements present, but lacks specific details)\n3rd=[3] (Unrelated to music, focuses on arm wrestling)\n4th=[5] (Unrelated activity in a parking lot)\n5th=[2] (Car wash scene unrelated to the query)\n</summary>\n</think>\n<answer> [4] > [1] > [3] > [5] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_yZLwR5EJ7Wc.mp4",
"activitynet/videos/v_Z-6dR4H2dns.mp4",
"activitynet/videos/v_eil9mWGJB8E.mp4",
"activitynet/videos/v_o0d1PXKkJqM.mp4",
"activitynet/videos/v_gHYcZYkrxxQ.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 thoroughly wash a car by spraying it with soap and then rinsing it with water, as the camera captures various close-up and wide-angle shots of the process.\"\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 scenes at a car wash, but does not appear to feature two men washing a car. Instead, it shows women in swimsuits performing car wash activities and various other non-car wash related images.\n\n[2] The series depicts a man clearing snow off a car. The activity focuses on snow removal rather than washing a car with soap and water.\n\n[3] This sequence accurately captures two men engaging in the process of spraying soap and rinsing a car with water. The shots include a mix of close-ups and wide angles, showing the car washing process detailed in the query.\n\n[4] The video includes varied scenes related to car washing, including soap application and rinsing, but the people involved are not two men throughout. The narrative and visuals do not closely match the query's specified actions.\n\n[5] The frames depict construction work, not car washing. The activity involves laying roofing materials rather than washing a car.\n</content>\n\n<contrast>\n- [3] beats [2] because it involves two men washing a car with soap and water, mirroring the query criteria, whereas [2] is about snow removal.\n- [3] beats [1] because [3] accurately depicts car washing with two men, whereas [1] focuses on promotional or staged visuals that don't portray the car wash process.\n- [3] beats [4] because it clearly shows the specified activity of two men washing a car, whereas [4] includes car wash elements but not with the exact query criteria.\n- [3] beats [5] because it is relevant to car washing, whereas [5] shows construction work, unrelated to the query.\n- [1] beats [4] based on its closer proximity to a car wash setting despite the presence of other elements, whereas [4] has fragmented narratives.\n- [1] beats [2] as it has more elements related to a car wash setting, whereas [2] is specifically about snow removal not washing.\n- [4] beats [2] because it includes some scenes of car cleaning processes, albeit not meeting all criteria of the query, while [2] does not match at all.\n- [2] beats [5] because it involves a vehicle even if contextually different, whereas [5] focuses on construction work unrelated to cars.\n</contrast>\n\n<summary>\n1st=[3] (thorough car wash, two men, various detailed shots)\n2nd=[1] (car wash setting, less precise to query)\n3rd=[4] (car cleaning elements, lacks specific query action)\n4th=[2] (unrelated snow removal)\n5th=[5] (construction work, not car-related)\n</summary>\n</think>\n<answer> [3] > [1] > [4] > [2] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Z-6dR4H2dns.mp4",
"activitynet/videos/v_xgIxKDm9M38.mp4",
"activitynet/videos/v_QJmoA3byOzo.mp4",
"activitynet/videos/v_PZjWkB_q2lE.mp4",
"activitynet/videos/v_r7E7cVoJ0JE.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 artist paints a vibrant seascape, transforming a blank canvas into a colorful depiction of mountains, sky, and ocean.\"\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 shows people chopping wood with axes in an outdoor setting. There is no painting or artistic creation involved.\n\n[2] The video features a person painting what appears to be abstract flowers on a canvas. The focus is on the painting process, but the subject is not a seascape.\n\n[3] This sequence involves a person painting a landscape featuring mountains, sky, and ocean. It appears to be a step-by-step creation of a seascape painting, aligning closely with the query.\n\n[4] The video depicts people on a beach playing with sand. There is no painting or depiction of a seascape on canvas.\n\n[5] This sequence shows a person applying sealant around various fixtures, likely involving some maintenance work. There is no artistic activity or painting depicted.\n</content>\n\n<contrast>\n- [3] matches the query as it depicts the creation of a seascape painting, whereas [2] features abstract flowers, not a seascape.\n- [2] shows a painting process, while [4] involves sand activities, not related to painting.\n- [4] is further from the query as it shows beach scenes and no painting.\n- [5] involves maintenance work which is unrelated to any artistic creation.\n- [1] involves chopping wood, not related to painting, thus less relevant than [2] and [3].\n</contrast>\n\n<summary>\n1st=[3] (depicts a seascape painting)\n2nd=[2] (involves painting, but not a seascape)\n3rd=[4] (unrelated to painting, involves beach activities)\n4th=[5] (involves maintenance, not related to art)\n5th=[1] (chopping wood, no relation to painting)\n</summary>\n</think>\n<answer> [3] > [2] > [4] > [5] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_yiFKPKXevOU.mp4",
"activitynet/videos/v_fRUoSevi63M.mp4",
"activitynet/videos/v_gc04Okw4DTs.mp4",
"activitynet/videos/v_oZgyW8Yca_M.mp4",
"activitynet/videos/v_PzAjZbCEbLU.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 stands on a diving board, jumps, and dives into the water.\"\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 captures a woman standing on a diving board, preparing to jump, and then diving into the water. It smoothly follows her movements, clearly showing the actions of standing, jumping, and diving.\n[2] The images depict a news segment about wakeboarding, showing a person wakeboarding with some close-ups and additional reporting, unrelated to diving.\n[3] This sequence shows underwater activities, with individuals swimming and using flippers, along with mermaid tails, focusing on swimming techniques rather than diving.\n[4] The footage centers on swimming techniques, consisting of instruction on arm and leg movements underwater, but does not involve diving from a board.\n[5] The images show a gymnast performing on a balance beam, displaying a different athletic routine inside a gym, not related to diving.\n\n</content>\n\n<contrast>\n- [1] features the complete action of diving from a board, whereas [2] focuses on wakeboarding.\n- [1] shows a diving sequence, whereas [3] involves swimming and underwater movements.\n- [1] is centered on diving into a pool, whereas [4] illustrates swimming techniques.\n- [1] showcases diving actions, whereas [5] presents gymnastics on a balance beam.\n</contrast>\n\n<summary>\n1st=[1] (Directly matches the query by showing the woman diving into water)\n2nd=[3] (Relates to water activities, although focused on swimming)\n3rd=[4] (Shows swimming techniques, some relation to water sports)\n4th=[5] (Features athletic performance, not in water)\n5th=[2] (Completely unrelated to diving, focused on wakeboarding)\n</summary>\n</think>\n<answer> [1] > [3] > [4] > [5] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_JNr0oI927ng.mp4",
"activitynet/videos/v_wdsZn1Yu_5M.mp4",
"activitynet/videos/v_XC6tvSBS0PA.mp4",
"activitynet/videos/v_kFmOtf_Ew1w.mp4",
"activitynet/videos/v_wlAnWPxnQ8k.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 prepares herself by having her hair brushed while occasionally looking at her phone and 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 images show people decorating a Christmas tree. There's no focus on a girl preparing herself or using a phone.\n\n[2] The images show a grooming session of a poodle. It does not involve a human girl or indicate any preparation or use of a phone.\n\n[3] The images display a tutorial on combing and blow-drying hair. The person is engaged in hair grooming and there's no visible phone or indication of preparing/personal grooming with phone interaction.\n\n[4] The images capture a scene with a woman at a computer, wearing curlers, and interacting with a second person. There's no brushing or looking at a phone while preparing.\n\n[5] The images depict a girl having her hair brushed while occasionally looking at a phone. It fits the query, showing the preparation action and phone interaction.\n</content>\n\n<contrast>\n- [5] matches the query as it involves hair brushing and phone usage, whereas [1], [2], [3], and [4] do not show the required actions.\n- [3] contains some hair brushing content but lacks phone interaction or a preparation scene involving a girl.\n- [4] involves curlers but no hair brushing combined with phone usage.\n- [1] and [2] are unrelated to the concept of personal preparation or phone interaction.\n</contrast>\n\n<summary>\n1st=[5] (matches query with brushing and phone use)\n2nd=[3] (some hair grooming but no phone or preparation)\n3rd=[4] (curlers present, no brushing with phone use)\n4th=[1] (decorating unrelated to personal grooming)\n5th=[2] (dog grooming unrelated to query)\n</summary>\n</think>\n<answer> [5] > [3] > [4] > [1] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_EeUkryzfwYk.mp4",
"activitynet/videos/v_W04glwJ_IYI.mp4",
"activitynet/videos/v_padyJHC5Y5Y.mp4",
"activitynet/videos/v_NS8qkB39x9U.mp4",
"activitynet/videos/v_gU67yCVXHnQ.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 thoroughly brushes a horse while another horse occasionally appears in the background.\"\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 a young girl thoroughly brushing a horse in an outdoor setting. Another horse occasionally appears in the background, matching the query closely.\n\n[2] This video shows a person brushing a horse indoors. While brushing is evident, the setting does not show an outdoor environment or another horse appearing in the background.\n\n[3] This video features a person riding a horse in an outdoor paddock, unrelated to grooming activities.\n\n[4] The sequence involves children on bumper cars in an amusement park, irrelevant to the query.\n\n[5] Similar to [4], this involves bumper cars at an amusement park, not related to horse grooming.\n</content>\n\n<contrast>\n- [1] is most relevant because it shows a young girl brushing a horse outdoors with another horse appearing in the background, fitting the query.\n- [2] involves brushing a horse but lacks the outdoor element and the presence of another horse, whereas [1] includes both.\n- [3] is unrelated as it shows horse riding instead of grooming, while [1] features horse grooming as specified in the query.\n- [4] and [5] involve bumper cars, completely unrelated to horse grooming, unlike [1].\n</contrast>\n\n<summary>\n1st=[1] (matches the query, showing a girl brushing a horse outdoors with another horse occasionally in the background)\n2nd=[2] (depicts brushing but lacks an outdoor setting and presence of another horse)\n3rd=[3] (shows horse riding, not grooming)\n4th=[4] (bumper cars, irrelevant to the query)\n5th=[5] (bumper cars, irrelevant to the query)\n</summary>\n</think>\n<answer> [1] > [2] > [3] > [4] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_YrS64TBX798.mp4",
"activitynet/videos/v_0PbmXOVIwGc.mp4",
"activitynet/videos/v_PnY3FT_QqWo.mp4",
"activitynet/videos/v_SsW-IeF0bzE.mp4",
"activitynet/videos/v_EZdnNs-5tlo.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 demonstrates how to use a snow blower to clear a path, periodically addressing the camera as she works.\"\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 someone using an iron and then snowboarding. This does not relate to snow blowing.\n[2] A person is mowing a lawn with a child on their back. This is unrelated to snow blowing.\n[3] The video features snow-covered mountains and skiing activities. There is no snowblower or demonstration involved.\n[4] This video appears to be an instructional guide on snow shoveling, including warnings and preparation tips. There is no snow blower demonstrated; the focus is on shoveling.\n[5] A woman is using a snow blower to clear a path of snow, and she appears to address the camera at times, related directly to the query.\n</content>\n\n<contrast>\n- [5] matches the query as it shows a woman using a snow blower and addressing the camera, whereas [1] shows snowboarding.\n- [5] involves snow removal with a snow blower, fulfilling the task in the query, whereas [2] is about mowing a lawn.\n- [4] relates to snow clearing but focuses on shoveling, not snow blowers, unlike [5].\n- [3] features snow activities but not snow clearing, unlike [5].\n- [5] has a clear demonstration of equipment use that matches the query, in contrast to the irrelevant content in [1], [2], and [3].\n</contrast>\n\n<summary>\n1st=[5] (demonstrates snow blowing and addresses the camera)\n2nd=[4] (related to snow removal but involves shoveling, not snow blowing)\n3rd=[3] (features snow-related activities, but not snow blowing)\n4th=[1] (involves snow but in a recreational context, not blowing)\n5th=[2] (unrelated activity of mowing a lawn)\n</summary>\n</think>\n<answer> [5] > [4] > [3] > [1] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_a6Wwa7qvlrA.mp4",
"activitynet/videos/v_kyMsxlpBjlY.mp4",
"activitynet/videos/v_Lr0Vm9zBXKE.mp4",
"activitynet/videos/v_kuyuhXuMPX4.mp4",
"activitynet/videos/v_jml_hnhjltg.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: \"This video showcases a lively beach scene with people building sandcastles, interviews with participants and a judge, and views of the event's setup and atmosphere.\"\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 beach volleyball game with players actively engaging in playing and serving the ball. It focuses on the sport rather than sandcastles or interviews.\n[2] This video features beach volleyball played in a large stadium with a crowd, cheerleaders, and festive activities. The focus is on the volleyball match and audience.\n[3] The video is about basketball drills inside a gym, which is unrelated to the query of a beach event with sandcastles and interviews.\n[4] This video perfectly matches the query. It shows people building sandcastles on the beach, includes interviews with participants, and presents the setup of a sandcastle event.\n[5] The video showcases a beach soccer event with children playing, a crowded beach atmosphere, and some small interviews, but the main focus is not on sandcastles.\n</content>\n\n<contrast>\n- [4] is ranked highest because it directly showcases people building sandcastles and includes interviews, closely fitting the query.\n- [5] is ranked second as it features a beach event with a lively atmosphere, though its focus is on soccer rather than sandcastles.\n- [1] is ranked third as it has a beach setting but focuses on volleyball, not sandcastle building or interviews.\n- [2] is ranked fourth because it involves a beach-related event but focuses on volleyball in a stadium, not sandcastle building.\n- [3] is ranked last as it is unrelated to beach activities, focusing instead on indoor basketball drills.\n</contrast>\n\n<summary>\n1st=[4] (perfect match with sandcastles, interviews, and event setup)\n2nd=[5] (beach atmosphere; some interviews but focus on soccer)\n3rd=[1] (beach setting with volleyball focus; lacks sandcastle element)\n4th=[2] (beach stadium volleyball; lacks sandcastles or interviews)\n5th=[3] (indoor basketball drills, unrelated to query)\n</summary>\n</think>\n<answer> [4] > [5] > [1] > [2] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_SLPD1U5iDjY.mp4",
"activitynet/videos/v_LE05u6TR9MI.mp4",
"activitynet/videos/v_DmTI8y6Anvo.mp4",
"activitynet/videos/v_gYARGADLcmQ.mp4",
"activitynet/videos/v_Mzojo2EeWu8.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 demonstrates a step-by-step nail art process, including applying polish with a sponge, using tape for clean edges, and finishing with a clear coat and cleanup.\"\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 preparing and chopping vegetables, specifically bell peppers and cucumbers. The actions shown include cutting and placing vegetables into a bowl, unrelated to nail art.\n\n[2] This sequence primarily focuses on nail art with steps showing the application of different colored nail polish and designs. It includes visuals of nails prepped and painted, but doesn't clearly demonstrate the use of a sponge, tape, or specific cleanup.\n\n[3] This sequence illustrates a process of making a dessert, involving mixing ingredients and baking. There is no visible connection to nail art or manicure techniques.\n\n[4] This video clearly demonstrates a step-by-step nail art process. It features the use of a sponge for gradients, tape for clean edges, multiple colors of nail polish, and a clear coat. The process closely aligns with the query.\n\n[5] This sequence shows a person interacting with a cat, including trimming the cat’s claws. There are no visible nail art techniques demonstrated here.\n</content>\n\n<contrast>\n- [4] is higher ranked than [2] because it clearly shows the use of tape and a sponge, matching the detailed steps in the query.\n- [2] beats [1] because it focuses on nails, though without all specified techniques; [1] is unrelated to nail art.\n- [1] is ahead of [3] because its manual dexterity aspect (cutting) is closer to nail art than baking.\n- [3] is ranked above [5] as it involves a creation process, whereas [5] involves a cat's manicure, unrelated to nail art processes involving polish and design.\n</contrast>\n\n<summary>\n1st=[4] (shows detailed nail art process with sponge, tape)\n2nd=[2] (focus on nail art but missing some techniques)\n3rd=[1] (manual dexterity closer to nail art)\n4th=[3] (creation process but unrelated task)\n5th=[5] (unrelated activity with a cat)\n</summary>\n</think>\n<answer> [4] > [2] > [1] > [3] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_tnavaGow7BI.mp4",
"activitynet/videos/v_EfQtRtIFGVo.mp4",
"activitynet/videos/v_YS3Q47uz3HI.mp4",
"activitynet/videos/v_TJFl99tLNWk.mp4",
"activitynet/videos/v_2tpwfPdSEVo.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 in athletic attire demonstrates and explains the proper technique for doing crunches while lying on a yoga mat.\"\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 appears to show interviews and a group exercise session. People are wearing athletic attire, but there's no demonstration or explanation of crunches on a yoga mat.\n\n[2] This sequence showcases a woman in athletic attire lying on a yoga mat. She demonstrates and likely explains the technique for doing crunches. The frames clearly follow a step-by-step process of the exercise.\n\n[3] A woman in athletic attire performs exercises on a yoga mat, also demonstrating movements similar to crunches, though the emphasis is slightly different compared to the traditional form.\n\n[4] The setting involves outdoor activities like rafting, unrelated to demonstrating crunches on a yoga mat.\n\n[5] The video captures a track and field environment, showing a woman in athletic attire but unrelated to a yoga mat or crunches demonstration.\n</content>\n\n<contrast>\n- [2] is the most relevant as it visually depicts the complete setup and execution of crunches on a yoga mat.\n- [3] includes exercises on a mat but focuses on a slightly different movement than the standard crunch, whereas [2] clearly shows the crunch exercise.\n- [1], [4], and [5] do not show relevant activities related to crunch exercises on a yoga mat, focusing on other athletic or outdoor activities instead.\n</contrast>\n\n<summary>\n1st=[2] (Direct demonstration of crunches on a yoga mat)\n2nd=[3] (Mat exercises shown but slightly different from standard crunches)\n3rd=[1] (Athletic activities unrelated to crunches)\n4th=[5] (Athletic activities unrelated to crunches)\n5th=[4] (Outdoor activities, not related to crunches)\n</summary>\n</think>\n<answer> [2] > [3] > [1] > [5] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_7OZHg_OOfxw.mp4",
"activitynet/videos/v_Ek2OvKMdvys.mp4",
"activitynet/videos/v_zNpILhjhFbQ.mp4",
"activitynet/videos/v_I9_eKiGZPxc.mp4",
"activitynet/videos/v_hS7VEMlJ4N0.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 teams play a game of shinty on a partially filled field, culminating in a winning moment that is celebrated by cheering 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] This sequence involves a field game with players in uniforms, resembling a lacrosse match. The footage includes text commentary such as \"The Sickest Lacrosse Play EVA!\" and \"How Did That Feel?\" There are spectators visible, but it's unclear if there is a winning moment celebrated by cheering spectators.\n\n[2] This video features a frisbee competition with dogs. No teams, shinty, or evident spectator cheering is present.\n\n[3] The sequence shows a tug-of-war event with two teams pulling on a rope. There is a crowd of spectators, but the sport does not relate to shinty or have a winning celebration shown.\n\n[4] This video contains a sport that resembles shinty or hurling, with players on a field using sticks to hit a ball. There are shots of the players in action and a crowd of spectators watching, with a celebratory scene in the final frames.\n\n[5] This sequence features a dog agility or frisbee event. There are spectators present, but no relation to shinty or team play is visible.\n</content>\n\n<contrast>\n- [4] showcases a shinty-style game with active player engagement and cheering spectators in the end, whereas [1] predominantly involves lacrosse with text overlays and no clear winning moment.\n- [1] involves human participants in a field game, whereas [2] & [5] are dog-centric events with no relevance to team sports like shinty.\n- [3] features a tug-of-war, which involves teamwork and spectators but does not sufficiently match the shinty game criteria.\n- [2] and [5] focus on dog agility/frisbee, not on shinty or a similar team sport.\n</contrast>\n\n<summary>\n1st=[4] (Features a shinty-style sport with team play and cheering spectators)\n2nd=[1] (Lacrosse game with spectators, less matching shinty specifics)\n3rd=[3] (Tug-of-war with teams and spectators, no shinty context)\n4th=[2] (Frisbee with dogs, lacks shinty and team elements)\n5th=[5] (Dog event unrelated to shinty or team sports)\n</summary>\n</think>\n<answer> [4] > [1] > [3] > [2] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Gc1Mk5UyECQ.mp4",
"activitynet/videos/v_e__FVkfXB-U.mp4",
"activitynet/videos/v_T3ngvDyxYiE.mp4",
"activitynet/videos/v_Xv2-EY_zKNM.mp4",
"activitynet/videos/v_eLeGJBfqURA.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: \"Five serious-looking people, including two women and three men, play piano accordions together to create harmonious 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] This sequence shows five individuals seated in a spacious room, all playing piano accordions. There are three men and two women, aligning with the query description. They appear serious and are likely engaged in playing harmonious music together.\n\n[2] The scenes depict a dynamic performance involving multiple performers in a well-lit stage environment, with activities such as marching and jumping rope. There is no evidence of piano accordions or a configuration of two women and three men playing instruments.\n\n[3] The frames are black-and-white and exhibit two young men in various everyday settings, including a barber shop and sitting indoors. The scenes contain text overlays indicating a narrative but do not depict the musical performance described in the query.\n\n[4] This video shows soldiers engaged in maintenance tasks, with a focus on boot cleaning and polishing. There is no musical performance or visible instruments, so it does not match the query.\n\n[5] These scenes focus on a musical duet, with a woman playing the piano and a man playing the saxophone. While musical, this does not feature piano accordions or the specific group configuration sought in the query.\n</content>\n\n<contrast>\n- [1] is a perfect match with the query because it shows five people, including two women and three men, seriously playing piano accordions together.\n- [2] is a performance but lacks instruments and the specific group arrangement described in the query.\n- [3] contains no musical performance, whereas [1] features one.\n- [4] focuses on non-musical tasks, whereas [1] directly depicts a musical performance.\n- [5] is musical but features only two performers and different instruments compared to the accurate portrayal in [1].\n</contrast>\n\n<summary>\n1st=[1] (shows the accurate group configuration and instruments described in the query)\n2nd=[5] (musical setting but incorrect instruments and fewer performers)\n3rd=[2] (dynamic performance but lacks the specific musical elements)\n4th=[3] (no musical performance; focuses on unrelated activities)\n5th=[4] (non-musical tasks; unrelated to the query)\n</summary>\n</think>\n<answer> [1] > [5] > [2] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_rBgMeunuviE.mp4",
"activitynet/videos/v_HysxcMFrTkc.mp4",
"activitynet/videos/v_d14twk9O2Vw.mp4",
"activitynet/videos/v_MnzVdvbvnMc.mp4",
"activitynet/videos/v_RVOBQBamjDA.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 paddle a canoe across a scenic lake, passing ducks, rocks, and tall grass before arriving at the shore.\"\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 shows a group of people in a blue inflatable raft navigating whitewater rapids. The frames capture moments of active paddling through turbulent waters. The setting is a fast-moving river rather than a calm scenic lake.\n\n[2] This sequence features people paddling a green canoe on a calm lake or river. It includes scenery such as rocks, tall grass, and passing animals, aligning with the query description of a scenic lake.\n\n[3] The images depict individuals in a canoe on a calm body of water. The sequence shifts to various non-related scenes like dining, socializing, and outdoor activities not related to paddling across a lake.\n\n[4] The sequence contains cartoon illustrations and graphics promoting roof cleaning services, with no connection to canoeing or a scenic lake environment.\n\n[5] This sequence shows an individual paddling a kayak on a lake. The setting suggests a larger body of calm water but lacks a group or scenic elements like ducks or tall grass as described in the query.\n</content>\n\n<contrast>\n- [2] matches the query best with scenes of people paddling a canoe in a scenic setting including rocks and vegetation, unlike [1], which shows whitewater rafting.\n- [3] features canoeing but includes unrelated content, whereas [5] focuses on a single paddler, not a group.\n- [5] shows a kayak on a lake, somewhat relevant but lacks the group and scenic context of [2].\n- [1] involves whitewater rafting, not matching the calm lake scenario of [2].\n- [4] is entirely unrelated, focusing on roof cleaning graphics.\n</contrast>\n\n<summary>\n1st=[2] (Matches the scenic canoeing query with appropriate visuals and context)\n2nd=[5] (Involves water paddling but lacks group and scenery details)\n3rd=[1] (Rafting, not canoeing, in rapids; action-oriented)\n4th=[3] (Partial canoeing content mixed with unrelated scenes)\n5th=[4] (Unrelated to canoeing or scenic lake settings)\n</summary>\n</think>\n<answer> [2] > [5] > [1] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_mNTO-xnsaTc.mp4",
"activitynet/videos/v_tl3wQEusmj8.mp4",
"activitynet/videos/v_wdecG9VSAl0.mp4",
"activitynet/videos/v_43gst-Mw43s.mp4",
"activitynet/videos/v_sQtT9sUyUwE.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 follows a group of people as they prepare for, learn about, and successfully experience an exhilarating white water rafting adventure with Outdoor Interlaken, concluding with celebration and contact information.\"\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 depict a relaxed river ride, not specifically a white water rafting adventure. There is no preparation, learning, or contact information depicted. The overall tone is calm and leisurely.\n\n[2] This sequence shows preparation for a rafting trip with equipment setup and safety talks, then transitions to scenes of adventurous rafting on rapids. There is a clear inclusion of celebration with the group holding a raft, and it ends with contact information for Outdoor Interlaken.\n\n[3] The video shows preparation stages with a group assembling, followed by rafting in rapids. However, there's no apparent celebration or contact information visible, and the branding focuses on \"Bali Adventure Tours.\"\n\n[4] The scenes include rafting activities but also feature unrelated elements like a device and mention of \"Drug Testing Outsourced,\" which distract from the focus on a cohesive rafting adventure narrative.\n\n[5] This displays individuals kayaking in intense water conditions with an emphasis on action rather than a group experience. There are no scenes of preparation, learning, or contact details related to a rafting company.\n</content>\n\n<contrast>\n- Item [2] offers a full narrative from preparation to celebration, with explicit contact information, whereas item [3] lacks clear celebration and contact details.\n- Item [2] provides comprehensive coverage tailored to the query, whereas item [4] introduces unrelated elements that detract from the adventure theme.\n- Item [2] has a coherent rafting adventure storyline, whereas item [1] shows a tranquil river ride lacking context relevant to the query.\n- Item [2] contrasts with [5] by presenting a group experience with preparation and contact elements, whereas [5] focuses solely on individual kayak action.\n</contrast>\n\n<summary>\n1st=[2] (Complete narrative with preparation, rafting, celebration, and contact info)\n2nd=[3] (Good coverage of preparation and rafting, lacks celebration/contact info)\n3rd=[4] (Has rafting scenes but includes unrelated content distracting from the theme)\n4th=[5] (Focuses on individual kayaking, missing group prep/celebration/contact)\n5th=[1] (Calm river ride not aligned with the dynamic rafting experience query)\n</summary>\n</think>\n<answer> [2] > [3] > [4] > [5] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Uk_VxavVjP0.mp4",
"activitynet/videos/v_n6k21NjvqXE.mp4",
"activitynet/videos/v_jIsGq431gBs.mp4",
"activitynet/videos/v_bi6C3kyP4mU.mp4",
"activitynet/videos/v_A1SuSeqDTEI.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 teen gymnast performs a pommel horse routine with spins and handstands, while being observed by a jury and onlookers.\"\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 male gymnast getting ready for parallel bars, performing handstands, but lacks a pommel horse routine or jury/observers.\n[2] A male gymnast is seen performing a parallel bars routine with spins and handstands, not on pommel horse nor visible jury/observers.\n[3] A female gymnast performs on a balance beam, unrelated to the pommel horse or query requirements.\n[4] This sequence features a male gymnast performing on a pommel horse, showing spins with a visible jury in the background, matching the query well.\n[5] A male gymnast is engaging in a pommel horse routine, including spins, with some visible onlookers, fitting the query, though less clearly than [4].\n</content>\n\n<contrast>\n- [4] is more relevant than [5] because it shows the pommel horse routine with a clear view of the jury.\n- [5] beats [1] as it features the pommel horse routine, whereas [1] is focused on parallel bars.\n- [1] is better than [2] since both are incorrect, but [1] is closer in apparatus type.\n- [2] ranks higher than [3] as parallel bars are closer in apparatus type to pommel horse than a balance beam.\n- [3] is least relevant as it involves a different gender, apparatus, and no visible observers/jury.\n</contrast>\n\n<summary>\n1st=[4] (clear pommel horse routine, spins, and jury presence)\n2nd=[5] (pommel horse routine, visible observers)\n3rd=[1] (gymnast focused on apparatus, though incorrect apparatus)\n4th=[2] (gymnast on parallel bars routine, action-related but wrong apparatus)\n5th=[3] (different apparatus and gender)\n</summary>\n</think>\n<answer> [4] > [5] > [1] > [2] > [3] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_ZT63nK4lexg.mp4",
"activitynet/videos/v_NfVL1o5HMEo.mp4",
"activitynet/videos/v_W6y6Vmk5edg.mp4",
"activitynet/videos/v_bmf7otW5TQo.mp4",
"activitynet/videos/v_AJ_QqCET1iY.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 explains and demonstrates the fundamentals of belaying in climbing through both direct narration and illustrative scenes outdoors.\"\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 a tutorial on dribbling in futsal, featuring two individuals demonstrating ball control. The frames show different techniques with captions in another language.\n[2] This sequence involves a man demonstrating how to install or replace flooring. The frames show hands-on work with wood and tile materials.\n[3] This video presents a man explaining and demonstrating belaying in climbing. The sequence includes outdoor scenes of climbing and spoken narration about belaying techniques, with text overlays indicating climbing terms.\n[4] A person is shown climbing a rock face in an outdoor setting. The sequence lacks any narration or belaying context.\n[5] This sequence shows several individuals involved in changing or repairing a wheel on a vehicle. The setting is outdoors, but unrelated to climbing or belaying.\n</content>\n\n<contrast>\n- [3] is directly related to the query as it shows a man discussing and demonstrating belaying in a climbing context, with relevant terms and actions illustrated.\n- [4], while involving climbing, lacks the instructional and belaying focus found in [3].\n- [1], [2], and [5] are unrelated to climbing or belaying. [1] focuses on futsal, [2] on flooring, and [5] on vehicle repair.\n</contrast>\n\n<summary>\n1st=[3] (directly matches the query with climbing and belaying demonstration and explanation)\n2nd=[4] (features climbing but lacks focus on belaying and instruction)\n3rd=[1] (unrelated, involves a sports tutorial)\n4th=[2] (unrelated, focuses on flooring installation)\n5th=[5] (unrelated, focused on vehicle maintenance)\n</summary>\n</think>\n<answer> [3] > [4] > [1] > [2] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_zGT6EplIOHY.mp4",
"activitynet/videos/v_ZIBWRRBft8g.mp4",
"activitynet/videos/v_JouywOCeGHI.mp4",
"activitynet/videos/v_OZpxc1X8Bsc.mp4",
"activitynet/videos/v_YoJxVl_hUWg.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 take turns jumping rope on a stage, coordinating their movements as one person holds the rope.\"\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 outside walking and jumping on bouncy stilts. It takes place in a park setting without coordinated movements on a stage. \n[2] A gym-like setting with various people standing around, some interactions, but not jumping rope or on a stage. \n[3] A group of children performing a dance on stage, but no jumping rope is involved. It's a coordinated stage performance.\n[4] People are taking turns jumping rope on a stage. There is coordination and one person holding the rope while others jump, consistent with the query.\n[5] A performance on a stage involving a slackline act, with coordinated movements. However, it does not involve jumping rope.\n</content>\n\n<contrast>\n- [4] precisely matches the query with people taking turns jumping rope on a stage, whereas [5] involves slackline acts, not jumping rope.\n- [3] features a group performance on stage but lacks the element of jumping rope, which [4] has.\n- [1] and [2] do not occur on a stage or involve jumping rope, making them less relevant compared to [4].\n- [5] shows a coordinated performance on stage, but it's slacklining, not rope jumping as in [4].\n</contrast>\n\n<summary>\n1st=[4] because it shows people taking turns jumping rope on a stage, aligning perfectly with the query.\n2nd=[5] because it has a stage performance, though it's slacklining instead of jumping rope.\n3rd=[3] because it shows a coordinated performance on stage but involves dancing, not jumping rope.\n4th=[1] because it has jumping activity, though not on a stage and doesn't involve rope.\n5th=[2] because it neither features a stage setting nor jumping rope.\n</summary>\n</think>\n<answer> [4] > [5] > [3] > [1] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_qxlVkDpdyOA.mp4",
"activitynet/videos/v_bWYIL4KG8kM.mp4",
"activitynet/videos/v_JxEnKQ7iXUI.mp4",
"activitynet/videos/v_fxEcsM0EaA4.mp4",
"activitynet/videos/v_oeFUzyNgxQU.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 dressed as pirates and in various costumes paddle canoes down a river, shooting flames into the air and celebrating 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] This video shows people in costumes, including pirate attire, paddling canoes down a river. There are visible celebrations and smoke indicating some pyrotechnic activities. This is directly aligned with the query.\n\n[2] This video depicts a farmers market with cooking activities, showcasing ingredients and people involved in culinary tasks. There are no visible elements related to paddling, costumes, or pyrotechnics.\n\n[3] This video involves a river rafting activity featuring people wearing helmets and life jackets. It lacks costumes and pyrotechnic elements as described in the query.\n\n[4] This video contains footage of a large outdoor event with many people, some of whom are near canoes on the river. However, there are no visible costumes or pyrotechnic elements.\n\n[5] This video features individuals rafting down a river in tubes. It focuses more on the rafting experience without costumes or pyrotechnic celebrations.\n\n</content>\n<contrast>\n- [1] beats [2] because it features costumes and river activities, whereas [2] is a cooking event.\n- [1] beats [3] because it shows costumes and celebrations, whereas [3] is purely rafting.\n- [1] beats [4] as it includes themed costumes, unlike [4], which lacks evident celebratory costumes.\n- [1] beats [5] for having costumes and celebratory elements, unlike the simple tubing in [5].\n- [4] beats [3] and [5] because it shows a crowd near canoes, suggesting a river event, whereas [3] and [5] lack costumes or celebration.\n- [3] is preferred over [5] because of the closer resemblance to a group river activity, even though both lack costumes and celebrations.\n- [2] is the least relevant as it is entirely unrelated to a river activity or costumes.\n\n</contrast>\n<summary>\n1st=[1] (direct match with costumes, river, and celebration)\n2nd=[4] (participants near canoes; suggests a themed event)\n3rd=[3] (river activity but without costumes)\n4th=[5] (river tubing without costumes or celebrations)\n5th=[2] (cooking event completely unrelated)\n\n</summary>\n</think>\n<answer> [1] > [4] > [3] > [5] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_p4YTDxTASBI.mp4",
"activitynet/videos/v_CIQ-mnURg9E.mp4",
"activitynet/videos/v_I9_eKiGZPxc.mp4",
"activitynet/videos/v_GLHsjvmOXRs.mp4",
"activitynet/videos/v_b_v5uGSheEw.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 smiles at the camera while people sail and spin in the water nearby.\"\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 shows people in bumper cars, enjoying a fun ride at an amusement park. In the final frame, a little girl is visible, but the environment is not water-related.\n\n[2] The frames depict individuals canoeing on a lake with marked lanes. There is no presence of a woman smiling at the camera.\n\n[3] A woman is seen smiling and engaging with the camera, followed by scenes of rafting with others paddling in a river. This closely aligns with the query’s theme of water activities and a woman smiling.\n\n[4] This video captures people rafting in a natural river setting, with various frames showing them paddling and navigating the water. There is no clear shot of a woman smiling at the camera.\n\n[5] A woman is seen smiling at the camera while others are windsurfing and sailing on the water nearby. This video directly matches the query criteria, showing both a woman and sailing activities.\n</content>\n\n<contrast>\n- [5] is ranked higher than [3] because it clearly shows both the smiling woman and the nearby sailing activities, perfectly matching the query.\n- [3] is preferred over [4] because it contains a smiling woman and water activity, while [4] lacks the smiling woman.\n- [4] beats [2] since it shows water activities, whereas [2] is focused on competitive canoeing and lacks the smiling woman.\n- [2] is ranked higher than [1] because [2] involves water activities (canoeing), whereas [1] only shows amusement park bumper cars and lacks the water aspect.\n</contrast>\n\n<summary>\n1st=[5] (clear smiling woman and sailing activity match)\n2nd=[3] (smiling woman with rafting, close relevance)\n3rd=[4] (rafting visible, lacks smiling woman)\n4th=[2] (canoeing in water, lacks smiling woman)\n5th=[1] (bumper cars, not water-related)\n</summary>\n</think>\n<answer> [5] > [3] > [4] > [2] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_SsW-IeF0bzE.mp4",
"activitynet/videos/v_8dXbbJWFEJo.mp4",
"activitynet/videos/v_I9_eKiGZPxc.mp4",
"activitynet/videos/v_6EN1q4TmPtI.mp4",
"activitynet/videos/v_AhWkUdag6NA.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 repeatedly swings across the monkey bars at a playground, demonstrating her playful energy and determination.\"\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 people on horseback traversing various terrains, including water and trails, indicating outdoor recreational activities. There are no monkey bars or playground scenes.\n\n[2] A girl is using a swing at a playground, suggesting playful activities, but lacks interaction with monkey bars.\n\n[3] Displays a playground slide being used by small animals and people. There are no monkey bars or children interacting with them.\n\n[4] Features a young girl actively swinging across monkey bars at a playground, clearly displaying energy and determination as described in the query.\n\n[5] Shows an individual participating in a bungee jump off a high platform. This setting and activity are unrelated to playground monkey bars.\n</content>\n\n<contrast>\n- [4] has the girl swinging across monkey bars, fitting the query directly, whereas [2] only shows swinging on a swing, not monkey bars.\n- [3] involves playground equipment but no monkey bars, whereas [4] shows monkey bars being used prominently.\n- [1] features horse riding, lacking any playground or monkey bars activity, whereas [4] directly matches the query.\n- [5] shows bungee jumping, a completely different activity, whereas [4] is perfectly aligned with the query.\n</contrast>\n\n<summary>\n1st=[4] (perfectly matches the query with monkey bars activity)\n2nd=[2] (playground setting, lacks monkey bar interaction)\n3rd=[3] (playground setting but unrelated equipment)\n4th=[1] (completely unrelated to playground or monkey bars)\n5th=[5] (entirely different activity, no playground or monkey bars)\n</summary>\n</think>\n<answer> [4] > [2] > [3] > [1] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_zxzBoK5t0XY.mp4",
"activitynet/videos/v_rZmNsUX-7SU.mp4",
"activitynet/videos/v_yVrru7t3478.mp4",
"activitynet/videos/v_pKV8lrvbX6Q.mp4",
"activitynet/videos/v_VZ2CPdIPa30.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: \"Jolie Glassman demonstrates proper punching and kicking techniques in a boxing gym as part of a Livestrong video on boxing.\"\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 in martial arts attire, performing various moves such as spins and kicks in what appears to be a gym-like setting, with visible punching bags.\n[2] This video features a person demonstrating boxing techniques in a gym setting, visible punching bag, and includes the Livestrong.com branding, suggesting boxing instruction content.\n[3] The video takes place in a biomechanics lab, showing individuals with sensors performing movements likely related to a study, not related to boxing or gym activities.\n[4] This video displays two individuals engaged in arm wrestling at a table, with instructions, unrelated to boxing or gym activities.\n[5] The video features a badminton match taking place in an indoor court, unrelated to boxing or gym demonstrations.\n\n</content>\n\n<contrast>\n- [2] is highly related because it shows boxing techniques and has Livestrong branding, while [1] shows martial arts movements which are less specific to boxing technique.\n- [3] lacks relevance as it shows a biomechanics lab without boxing content, whereas [1] at least features a punching and kicking environment.\n- [4] focuses on arm wrestling, unrelated to the query, whereas [1] involves martial arts and may include some aspect of punching.\n- [5] is about badminton, not related to boxing or gym demonstrations, whereas [1] involves a gym and punching bags, though not specific to boxing techniques.\n</contrast>\n\n<summary>\n1st=[2] (Shows boxing techniques with Livestrong branding)\n2nd=[1] (Martial arts in a gym with punching bags)\n3rd=[3] (Biomechanics lab, irrelevant to boxing)\n4th=[4] (Arm wrestling, not boxing-related)\n5th=[5] (Badminton match, unrelated context)\n</summary>\n</think>\n<answer> [2] > [1] > [3] > [4] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_tilmCzSf1oc.mp4",
"activitynet/videos/v_Y6UKk3t8Hj8.mp4",
"activitynet/videos/v_MINjkEBIIkA.mp4",
"activitynet/videos/v_w6kBE7BHHeU.mp4",
"activitynet/videos/v_U-N92yQynQA.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: \"This video demonstrates the process of changing a spare tire by removing the first tire and replacing it.\"\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 series of military scenes and shoe cleaning activities, unrelated to tire changing.\n\n[2] This video shows the use of a tire repair kit but not the process of removing and replacing a tire. It includes images related to cars and repair tools.\n\n[3] The video illustrates a detailed process of assembling a tire with visible tools and hands-on actions, indicating tire handling activities but not clearly showing tire removal/replacement.\n\n[4] This sequence demonstrates the process of changing a tire. It clearly shows removing the old tire and replacing it with a spare.\n\n[5] The video shows people clearing snow off cars, which is unrelated to changing tires.\n\n</content>\n\n<contrast>\n- [4] beats [3] because it clearly shows the action of changing tires whereas [3] only demonstrates tire assembly.\n- [3] beats [2] because it involves hands-on tire handling whereas [2] shows using a repair kit without tire removal/replacement.\n- [2] beats [5] because it is more related to car and tire maintenance whereas [5] focuses on snow removal from cars.\n- [5] and [1] are both unrelated to tire changing, but [5] involves cars more directly than [1] which focuses on military activities.\n\n</contrast>\n\n<summary>\n1st=[4] (Directly shows replacing a tire)\n2nd=[3] (Focuses on tire assembly and handling)\n3rd=[2] (Shows using a repair kit, but not tire replacement)\n4th=[5] (Involves cars but only cleaning snow)\n5th=[1] (Unrelated to tires and cars)\n\n</summary>\n</think>\n<answer> [4] > [3] > [2] > [5] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_MnzVdvbvnMc.mp4",
"activitynet/videos/v_98MoyGZKHXc.mp4",
"activitynet/videos/v_4r0qYLCucEI.mp4",
"activitynet/videos/v_Q7eK0xT6VSQ.mp4",
"activitynet/videos/v_3UbQ0UDmbbw.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 washes clothes by hand using water from buckets and a sink, as an unplugged washing machine sits unused nearby.\"\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 in an apron at a sink, possibly washing something, with assistance from a young girl. There's no visible washing machine or explicit clothing washing activity.\n[2] The video features a man using a knife sharpener outdoors. It doesn't relate to clothes washing or show a washing machine.\n[3] A woman and young girl are at a kitchen sink washing dishes. There’s no indication of washing clothes or a nearby washing machine.\n[4] A woman is using water from various buckets and sinks to wash clothes. An unused washing machine is visible nearby, matching the query description.\n[5] The video displays a person cleaning a bathroom sink and toilet, unrelated to hand washing clothes with a visible washing machine.\n</content>\n<contrast>\n- [4] matches the query by showing hand washing of clothes with an unused washing machine, whereas [1] shows dishwashing.\n- [1] features sink use but relates to dishwashing, whereas [2] focuses on knife sharpening.\n- [2] is irrelevant with outdoor knife sharpening, whereas [3] focuses on dishwashing indoors.\n- [3] involves dishwashing with no clothes or washing machine present, whereas [5] is focused on bathroom cleaning.\n- [5] is about cleaning a bathroom, lacking any clothes washing action, whereas [4] explicitly shows clothes washing.\n</contrast>\n<summary>\n1st=[4] (woman washing clothes by hand with unused washing machine present)\n2nd=[1] (domestic activity involving a sink, but unrelated to query)\n3rd=[3] (sink activity, but it's dishwashing)\n4th=[5] (involves cleaning, unrelated to clothes washing)\n5th=[2] (unrelated knife sharpening activity)\n</summary>\n</think>\n<answer> [4] > [1] > [3] > [5] > [2] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_QfUTGmlNO1s.mp4",
"activitynet/videos/v_28P5f3p_32g.mp4",
"activitynet/videos/v_XbSQ_7vh7yc.mp4",
"activitynet/videos/v_eaR8oEBlYWk.mp4",
"activitynet/videos/v_eMB9L-LDK1E.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 prepares a stuffed bread bowl sandwich by chopping vegetables, cooking meat, mixing ingredients, and assembling everything into a bread bowl, which is then wrapped and presented.\"\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 involves the preparation of a sandwich using bread, meat, and various ingredients. The process includes chopping vegetables, cooking meat, and assembling these elements into a bread bowl, finishing with wrapping the sandwich.\n\n[2] The sequence depicts making a soft dough, forming it into small balls, pressing with a fork, and baking them. It involves baking, not relevant to preparing a stuffed bread bowl sandwich.\n\n[3] The images show a drink preparation process involving cutting fruits, adding them to glasses, and mixing drinks. This is unrelated to the preparation of a stuffed bread bowl sandwich.\n\n[4] The sequence shows a person carving a pumpkin. This activity doesn't involve food preparation related to bread bowls or sandwiches.\n\n[5] This sequence involves peeling fruits or vegetables, but it does not proceed to include cooking, mixing, or assembling into a bread bowl, making it less relevant.\n</content>\n\n<contrast>\n- [1] beats [2] because it shows the preparation of a stuffed sandwich in a bread bowl, whereas [2] is about baking dough balls.\n- [1] is favored over [3] as it directly involves sandwich making, whereas [3] is about drink preparation.\n- [1] is preferred to [4] because it is about cooking a sandwich, while [4] shows pumpkin carving.\n- [1] is more relevant than [5] since it includes assembly and wrapping, while [5] only involves peeling.\n- [5] is better than [2], [3], and [4] because it involves some food preparation, relevant for sandwich making.\n- [2], [3], and [4] are less relevant, but [2] involves food preparation, thinly tying to the query, whereas [3] and [4] do not.\n- [3] involves drink preparation, while [4] is entirely unrelated to food preparation or assembly.\n</contrast>\n\n<summary>\n1st=[1] (Fully relevant: chopping, cooking, assembling, wrapping a bread bowl sandwich)\n2nd=[5] (Partially relevant: involves peeling, relevant in context of food prep)\n3rd=[2] (Slightly relevant: involves some food preparation)\n4th=[3] (Unrelated: drink preparation)\n5th=[4] (Unrelated: carving activity, not related to food prep)\n</summary>\n</think>\n<answer> [1] > [5] > [2] > [3] > [4] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_B3ZL18f1x4Y.mp4",
"activitynet/videos/v_uqd8A4iJ6Bo.mp4",
"activitynet/videos/v_6lIi6-Y6q10.mp4",
"activitynet/videos/v_z6g5QbIPatk.mp4",
"activitynet/videos/v_LW5lcM_rKBY.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 creates a pumpkin carving by first sketching a design on paper, then drawing and carving it onto a pumpkin, and finally displaying the illuminated result.\"\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 involves people discussing or exhibiting a product related to wood bowls and possibly some type of oil application. There is no pumpkin or carving activity related to the query.\n\n[2] This sequence displays a machine carving a jack-o'-lantern face into a pumpkin, starting with a blank pumpkin and ending with a lit jack-o'-lantern. This video involves carving and the illuminated result, but lacks a initial design sketching step on paper done by a girl, as the entire carving process is automated.\n\n[3] The sequence captures a person sharpening tools on a sharpening stone. There is no indication of pumpkin carving or sketching related to the query.\n\n[4] The video sequence shows a girl sitting at a table, going through the steps of carving a pumpkin: sketching a design, carving it, and displaying an illuminated pumpkin. This perfectly aligns with the query.\n\n[5] This video is unrelated, showing children in various settings, possibly at school or engaging in activities like drawing. The content is not relevant to pumpkin carving or creating.\n</content>\n\n<contrast>\n- [4] features a girl actively sketching and carving a pumpkin and shows the final illuminated product, whereas [2] uses an automated machine and lacks the personal sketching step.\n- [2] presents a coherent carving and lighting process but misses the initial design sketch step, whereas [4] includes all requested steps including sketching.\n- [3] and [5] do not include any elements of pumpkin carving, whereas [4] aligns perfectly with the query.\n- [1] focuses on wood products and lacks relevance to the pumpkin carving process.\n</contrast>\n\n<summary>\n1st=[4] (captures all parts of the query: sketching, carving, displaying)\n2nd=[2] (good carving process, missing sketching by a person)\n3rd=[1] (completely unrelated, at least involves product handling)\n4th=[3] (unrelated tool sharpening, no pumpkin element)\n5th=[5] (entirely irrelevant school and activity scenes)\n</summary>\n</think>\n<answer> [4] > [2] > [1] > [3] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_pn1XDTlNMso.mp4",
"activitynet/videos/v_pv6CNBFhgCU.mp4",
"activitynet/videos/v_Anvhv9FFv1Q.mp4",
"activitynet/videos/v_z6g5QbIPatk.mp4",
"activitynet/videos/v_mBZAFJ-Q6Mw.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 an ax to repeatedly cut a tree trunk into smaller pieces beside a dirt area.\"\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 demonstrating how to repair a PVC roof using adhesive tape and tools like scissors and a dry rag. This is an indoor repair task and involves no outdoor elements or cutting of tree wood.\n\n[2] The sequence depicts a person using an ax to chop a tree trunk into smaller logs within a dirt area, matching the query closely. The action and setting described in the query are both present here.\n\n[3] The frames illustrate a man using an ax to chop wood in an outdoor setting, featuring actions such as swinging and splitting wood. While it matches the query's action, the setting is less clearly defined as a dirt area.\n\n[4] This video involves gardening work, specifically placing mulch around a tree, with no ax usage or tree cutting involved. It focuses on plant care rather than wood chopping.\n\n[5] The frames depict an indoor construction process where workers are installing or working on materials like tiles. There is no relation to cutting wood or using an ax outside.\n</content>\n\n<contrast>\n- [2] directly shows a person using an ax in a dirt area to cut wood, perfectly aligning with the query because it involves the precise activity and setting described.\n- [3] also involves wood chopping with an ax but lacks the clear depiction of a dirt area present in [2].\n- [1], [4], and [5] are unrelated to the task and setting described in the query. [1] focuses on roof repair, [4] focuses on gardening, and [5] focuses on construction.\n</contrast>\n\n<summary>\n1st=[2] (precisely matches query: ax usage in a dirt area cutting wood logs)\n2nd=[3] (matches ax usage and wood cutting, but setting less defined)\n3rd=[1] (unrelated to ax or tree cutting; focuses on roof repair)\n4th=[4] (unrelated task; involves tree care, no ax usage)\n5th=[5] (unrelated to ax usage or any aspect of the query)\n</summary>\n</think>\n<answer> [2] > [3] > [1] > [4] > [5] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_VcyO9tZ3NkY.mp4",
"activitynet/videos/v_MsBqlgdzeag.mp4",
"activitynet/videos/v_2vy0dMXhlWI.mp4",
"activitynet/videos/v_YAjqqbSsFTg.mp4",
"activitynet/videos/v_XkWO7aqcC8Y.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 introduces a film, featuring interviews with actors, clips and stills from their work, culminating in a couple performing a tango routine on stage to enthusiastic applause.\"\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 person welding in a workshop, with protective gear and equipment visible. There is no context related to film introduction or a tango routine, making it irrelevant to the query.\n[2] Shows an acrobatic performance on a stage with athletes engaged in a tightrope walk, with an audience visible, but it does not contain film introduction or tango elements.\n[3] Features an aerobics class with multiple participants in a studio setting. There are no film introductions, interviews, or tango performances.\n[4] Contains various scenes from a dance show, including a couple performing a tango on a stage. Audience members are visible, and there are segments possibly showing interviews or introductions, matching the query closely.\n[5] Shows a contemporary dance performance on stage, featuring dancers in minimalistic attire. However, there are no interviews, film clips, or tango references associated with the query.\n</content>\n\n<contrast>\n- [4] beats others due to the presence of a tango performance and potential interviews, aligning with the query of film introduction and clips culminating in a tango.\n- [5] involves a dance performance which is relevant but lacks film introduction or tango specifics found in [4].\n- [2] involves performance but not related to films or tango.\n- [3] is an indoor aerobics session irrelevant to the film introduction or tango performance.\n- [1] involves welding, which does not relate to film or dance.\n</contrast>\n\n<summary>\n1st=[4] (Complete alignment with film introduction, tango performance, and audience interaction.)\n2nd=[5] (Dance performance without film context or specific tango focus.)\n3rd=[2] (Performance setting but unrelated to film or tango.)\n4th=[3] (Irrelevant aerobics session without film context.)\n5th=[1] (Unrelated welding activity with no connection to film or tango.)\n</summary>\n</think>\n<answer> [4] > [5] > [2] > [3] > [1] </answer>",
"role": "assistant"
}
] |
[
"activitynet/videos/v_Launtf-qjDM.mp4",
"activitynet/videos/v_oeFUzyNgxQU.mp4",
"activitynet/videos/v_TIfAkOBMf5A.mp4",
"activitynet/videos/v_Z3NtsxP6w-g.mp4",
"activitynet/videos/v_2SMmL6kIx-w.mp4"
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.