EVAThink: Evidence Sufficiency-Guided Training for Adaptive Thinking in Video Understanding

EVAThink. Evidence Sufficiency-Guided Training for Adaptive Thinking in Video Understanding. By Changti Wu and four others, from East China Normal University, the Zhongguancun Academy, and one other institution. Published on the tenth of November.

A vision-language model watching a video gets a fixed frame budget, and what it can do with a question depends on whether the frames it sampled carry enough evidence to answer it. When the evidence is sufficient, answering directly is the right move. When the observations are less informative, reasoning that connects the available cues may help instead. The difficulty is that the model has to choose between those two modes without being told which situation it is in.

EVAThink trains for that choice, guided by evidence sufficiency. The supervision has to come from somewhere, so the first component builds it. Recoverable Evidence Gap Mining constructs an evidence-dense view and an evidence-sparse view of a video which share the same question, the same answer, and an identical frame budget. It then keeps only the pairs where the sparse view yields lower direct-answer accuracy than the dense one, and yet benefits from explicit reasoning. Those pairs are precisely the cases where response mode matters, and they become the training signal for choosing it.

Training itself is Evidence-Guided Two-Stage Reinforcement Learning. The stages carry different rewards: the first guides the model toward evidence-based mode preferences, the second refines how efficient its responses are. A first-token advantage regularization runs alongside, to keep the model from collapsing onto one mode.

The curated training set is small, six thousand seven hundred samples. On that much, EVAThink consistently improves on its base model across video understanding benchmarks, both under a thirty-two frame budget and under a sixty-four frame one, and it outperforms strong baselines on average.

Adaptive thinking here is not a knob someone sets. It is learned from pairs built to isolate the one condition under which thinking earns its tokens: frames too sparse to answer from, but rich enough to reason over.