sy1998/video-oasis-reliance-study

Video-Oasis reliance user study

100 short videos (at most 180 s), each with one multiple-choice question and the responses of six models, i.e. 600 responses in total. Each response contains the model's full reasoning, the video time segments it cites, and its answer, together with a reliance-gap score (rg) that measures how much the model's answer depends on the segments it cites: around 0 means the answer does not depend on the cited evidence; larger values mean it does.

Download

All files are packed in video_oasis_reliance_study_v3.zip (1.4 GB). Unzip it to get README.md, study_sheet.md, study_key.csv, study_items.jsonl and the videos/ folder.

Models

• Three open-source video reasoning models that cite time segments, each prompted with its own released template: Video-Thinker, VideoChat-R1, VideoChat-R1.5.
• Two general-purpose models used zero-shot (original released weights): Qwen2.5-VL-7B-Instruct and Qwen3-VL-8B-Instruct. They are asked to first give the relevant time segment(s) as [start, end] in seconds, then a brief description of the visual evidence and their reasoning, then the option letter. The opening <think> tag is part of the prompt and is shown at the start of their responses.
• One model of ours: Qwen3-VL-8B-Instruct fine-tuned (supervised) to answer with the time segment it relies on; its responses contain only the cited segment(s) and the option letter.
• Greedy decoding for all models.

Questions and selection

• Questions are sampled from the official Video-Oasis question list (sub-benchmarks: ImplicitQA, LongVideoBench, MINERVA, MMR-V, MVBench, TVBench, VCR-Bench, VSI-Bench, Video-Holmes, Video-MME, egoschema), excluding RTV-Bench, one question per video, spread over the sub-benchmarks.
• A video is included only if all six responses cite at least one time segment and have an rg. Videos are not selected by rg or by answer correctness.

Files (inside the zip)

• study_sheet.md: per video, the question, options, and the six responses (Response A-F; model names hidden, order shuffled per video), each with the full reasoning, cited time segments and answer; no scores.
• study_key.csv: one row per response (video id, response letter, model, sub-benchmark, question uid, gold answer, model answer, rg, video file).
• study_items.jsonl: one JSON object per response with all fields: id, video_id, response, uid, benchmark, video, duration_s, question, options, gold_answer, model, prediction_process, cited_segments (intervals_s / points_s, in seconds), model_answer, rg.
• videos/: the 100 videos (file names normalized to ASCII, prefixed with the sub-benchmark).

Video sources and license

The videos are redistributed from the original benchmarks listed above, unmodified except for file names, solely to make this study reproducible. Copyright and license terms remain those of the original datasets and video sources; please consult the original benchmarks before any other use. The model outputs and scores in this repository are ours.