Video models learn to check the evidence before reasoning: EVAThink hits 63.5% average at 32 frames with 6.7K samples and about 62% shorter output than always reasoning
Synopsis
EVAThink turns whether a sampled set of frames is enough to answer directly into a training signal: for the same question and answer it builds evidence-dense and evidence-sparse views of a video and keeps only pairs where direct answering degrades on the sparse view but explicit reasoning recovers accuracy, then uses two-stage reinforcement learning to teach the model to switch between direct answering and reasoning. Trained on only 6.7K samples, it reaches 63.5% and 65.6% average across six video understanding benchmarks under fixed 32- and 64-frame budgets, both above the Qwen3-VL-8B-Instruct base; at 64 frames the adaptive mode beats direct-only and think-only by 2.3 and 3.6 percentage points and cuts output length by about 62% relative to think-only.
Interpretation
Under a fixed frame budget, a video understanding model no longer always answers directly or always reasons; it picks a response mode according to whether the sampled frames support an answer. Earlier adaptive-reasoning methods mostly decide whether to reason from initial-answer confidence or from a trade-off between answer quality and generation cost, leaving supervision across different evidence conditions less explored. Under fixed 32- and 64-frame budgets, EVAThink averages 63.5% and 65.6% across six video understanding benchmarks, 0.6 and 1.4 percentage points above Qwen3-VL-8B-Instruct, and improves on that base on every benchmark.
The training data come from contrasting evidence-dense and evidence-sparse views of the same question and answer rather than from annotated reasoning demonstrations. Only pairs where direct-answer accuracy drops on the sparse view but explicit reasoning raises it are retained, tying the evidence condition directly to the benefit of reasoning. From a 118.2K candidate pool, 1,766 high-confidence pairs are kept for Stage 1 and 4,179 for Stage 2, for 6.7K training samples in total, versus 83K for VideoAuto-R1.
Two-stage reinforcement learning first shapes response-mode preferences from evidence conditions and then shortens responses, with first-token advantage regularization preventing collapse onto a single mode. Stage 1 prefers direct answering on dense views and reasoning on sparse views; Stage 2 drops the view-dependent preference and adds a small bonus for correct direct answers. Removing the Stage 1 mode-preference reward lowers the average from 65.6% to 61.2% and raises average output from 213 to 392 tokens; removing the Stage 2 direct-answer bonus leaves accuracy comparable (65.3% versus 65.6%) but raises output to 465 tokens.
Perspective
The demonstrated setting is the Qwen3-VL-8B-Instruct base, six video understanding benchmarks under fixed 32- and 64-frame budgets with a fixed per-frame pixel budget, and 6.7K training samples; AdaThinkV and Video-FLAIR are reported under their own original evaluation protocols, and AdaThinkV allows up to 256 frames. A reader interested in other bases or real deployment sampling policies should watch for replication under those conditions.
The authors state that recoverability means improved answer accuracy, not recovery of unobserved visual information; evidence segments are annotated by a stronger model, the annotation only supplies candidate regions, and retention is ultimately decided by two filtering rounds based on rollout accuracy. In the ablation, omitting Stage 2 training gives 65.4% average, slightly below the full model, so that stage's gain shows up mainly as shorter output.