7b-multimodal-models-long-video-qa-multi-turn-agentic-exploration-harness CLM-0196For 7B multimodal models on long-video QA, a multi-turn agentic exploration harness (adaptive viewing speed, cross-turn textual notes, rewatch) lowers accuracy below plain uniform-frame single-pass answering unless the model is trained to use it, because the untrained model commits to an answer before gathering evidence; with cold-start SFT on verified exploration turns plus trajectory-level RL, the same harness beats the single-pass baseline.
Capability: Reasoning about time in video
Observed on
Qwen2.5-VL-7B-Instruct backbone; nine long-video and cross-domain multiple-choice benchmarks; harness with fixed per-turn visual token budget..
Sources
- Measured: untrained agentic variant averaged 33.1 vs 41.8 for the single-pass base model on four long-video benchmarks, and 34.3 vs 37.8 cross-domain; the trained model reached 48.0 and 45.0. Ablation also shows RL without SFT collapses. Single backbone, single harness design, so it does not separate harness-specific quirks from a general prompt-only limit.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- On multi-hop video questions written so no single frame or timestamp lookup suffices, single-turn MLLMs answer poorly (0.07-0.51 accuracy across 12 proprietary and open models), and wrapping the same models in an agentic loop helps only when the loop lets the agent choose what, where and at what granularity to look: frameworks giving adaptive fps/resolution/interval control raised accuracy and trajectory scores for nearly every model, while a fixed-schedule framework with weak perception tools (STAR) fell below the single-turn baseline for all of them.Reasoning about time in video · reviewed by AI
- For 4-8B multimodal models on long-video QA, RL training on questions authored from short scene segments but answered over the whole video improves benchmark accuracy more than RL on questions authored from the whole video, because segment-authored questions target details a whole-video view overlooks; including the segment timestamp in the training instruction adds a further gain by removing referential ambiguity, and is not needed at inference.Reasoning about time in video · reviewed by AI
- For 7-8B multimodal models answering multiple-choice long-video questions under a fixed 16-frame budget, selecting keyframes by embedding similarity to the query concatenated with each candidate answer beats query-only scoring, query-agnostic diversity selection, and uniform sampling across Qwen2-VL, LLaVA-Video and InternVL3.5, with the largest margins on the longest video splits — but per-category results still invert, including uniform sampling winning some categories.Reasoning about time in video · reviewed by AI
- For long-video QA under a fixed frame budget with a frozen 7B multimodal model, organizing frames into a question-agnostic hierarchical temporal tree and searching it coarse-to-fine (answer, then issue a search statement, then expand the relevant branch) raises accuracy over uniform sampling and over agent methods that score frames independently, with the gain largest on the longest-video benchmarks.Reasoning about time in video · reviewed by AI
- For deepfake video detection with 7-8B open multimodal backbones, splitting the analysis into four separately prompted perspective agents (texture, lighting, motion, physics) whose reports a judge agent reconciles beats the same backbone doing chain-of-thought or multi-turn self-questioning over the same four perspectives, and the gain is not explained by extra turns or repeated frame access.Getting more from several models than from one · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: An untrained 7B backbone placed in the same multi-turn exploration harness scoring at or above its single-pass uniform-sampling baseline on these benchmarks. Drafted stance toward handing-agent-context-editing-tools-delete-summarize-note-fold-prompt: supports -- Same pattern: context-management tools given by prompt alone hurt mid-size models and only help after fine-tuning on trajectories that use them. Proposed technique, not catalogued: trained adaptive-pacing video exploration agent.