EVAThink trains video models to reason when frame evidence is sparse
Original title: EVAThink: Evidence Sufficiency-Guided Training for Adaptive Thinking in Video Understanding
Using 6.7K curated samples, it improves the base model under both 32- and 64-frame budgets.
EVAThink is a training method for vision-language video understanding that chooses direct answers or extra reasoning based on whether sampled frames contain enough evidence. Recoverable Evidence Gap Mining builds dense and sparse views of the same question under a fixed frame budget, then two-stage reinforcement learning sets mode preference and response efficiency. The authors say it improves the base model and beats strong baselines on average; code, model, and datasets are on GitHub.
- Sparse views are kept when direct answers fail but explicit reasoning helps.
- First-token advantage regularization is used to limit mode collapse.
- Training uses only 6.7K curated samples.
- Results are reported for 32-frame and 64-frame budgets.
- AI models
- EVAThink
From the original article
1 min read at huggingface.co
Under a fixed frame budget, video understanding with vision-language models (VLMs) depends on whether the sampled frames provide sufficient evidence for the question. Sufficient frame evidence supports direct answering, whereas less informative observations may benefit from reasoning that connects the available cues.
By Changti Wu, Ruizhe Fan, Bin Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.