EVAThink Teaches Video Models When to Think and When to Answer

Introduction

A video vision-language model does not always need a long reasoning trace. If the sampled frames clearly contain the answer, direct generation may be enough. If important events are only partially visible, however, reasoning can help connect weak or scattered cues. EVAThink addresses this decision problem by training the model to relate its response mode to the sufficiency of available visual evidence.

How EVAThink works

• Mining a recoverable evidence gap. The method creates evidence-dense and evidence-sparse views of video content while keeping the question, answer, and frame budget identical. This setup isolates the effect of what the sampled frames reveal, rather than simply rewarding a larger visual input.
• Keeping samples where reasoning matters. A pair is retained when the sparse view performs worse on direct answering than the dense view, but explicit reasoning improves the sparse-view result. Such examples teach a more useful distinction: limited evidence does not automatically mean the question is impossible; it may mean that the model should spend effort connecting the cues that remain.
• Training with two-stage reinforcement learning. The first stage focuses on learning evidence-based preferences between direct answers and reasoning responses. The second stage refines response efficiency. First-token advantage regularization is added to reduce collapse toward always thinking or always answering directly. The reported approach therefore targets both accuracy and the cost of unnecessary reasoning.

Results and implications

EVAThink is trained on only 6.7K curated samples. According to the supplied material, it consistently improves over the base model across video understanding benchmarks with both 32- and 64-frame budgets, and surpasses strong baselines on average. The source does not provide the full benchmark table or task-level scores, so the strongest supported conclusion is that evidence-guided response selection is a promising training direction rather than a universal win on every individual task.

The broader contribution is a shift from fixed reasoning policies to conditional computation. Clear visual evidence can support a shorter answer, while incomplete but recoverable evidence can justify additional reasoning. EVAThink does not merely assign a fixed reasoning style to a question category; it uses the relationship between the question and the observed frames as the training signal. This idea could complement future systems that actively select frames, locate relevant evidence, or retrieve moments from long videos. Such combinations may let a model decide not only how long to think, but also what visual information it should inspect first.

Source: Hugging Face Daily Papers