EVAThink: Evidence Sufficiency-Guided Training for Adaptive Thinking in Video Understanding
Abstract
EVAThink trains vision-language models to adaptively choose between direct answering and reasoning based on whether sampled video frames provide sufficient evidence, improving video understanding under fixed frame budgets.
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. We introduce EVAThink, which enhances video understanding through evidence sufficiency-guided training for adaptive thinking. First, Recoverable Evidence Gap Mining constructs evidence-dense and evidence-sparse video views that share the same question and answer under an identical frame budget. It retains pairs where sparse views yield lower direct-answer accuracy than dense views but benefit from explicit reasoning, providing supervision for response-mode selection. Second, Evidence-Guided Two-Stage Reinforcement Learning uses stage-specific rewards to guide training with evidence-based mode preferences and then refine response efficiency, while first-token advantage regularization mitigates mode collapse. Trained on only 6.7K curated samples, EVAThink consistently improves upon the base model across video understanding benchmarks under both 32- and 64-frame budgets, while outperforming strong baselines on average. The code, model, and datasets are available at https://github.com/ChangtiWu/EVAThink{EVAThink}.
Community
EVAThink: Evidence Sufficiency-Guided Training for Adaptive Thinking in Video Understanding
When should a video VLM think, and when should it answer directly? We introduce EVAThink, which learns to make this choice based on the sufficiency of available visual evidence.
💡 Key Ideas
- Recoverable Evidence Gap Mining: Construct evidence-dense and evidence-sparse video pairs with the same question, answer, and frame budget, retaining pairs where reasoning improves accuracy on sparse views.
- Evidence-Guided Two-Stage RL: First learn evidence-based response-mode selection, then refine efficiency—with no SFT cold start.
- Stable training: First-token advantage regularization mitigates collapse into always thinking or always answering directly.
📊 Results
Trained on only 6.7K curated samples, EVAThink consistently improves upon the base model across video understanding benchmarks under both 32- and 64-frame budgets, while outperforming strong baselines on average.
🚀 Why It Matters
Visual evidence provides a useful signal for allocating reasoning effort, enabling more accurate and efficient video understanding.
💻 Code, model & datasets: EVAThink
Get this paper in your agent:
hf papers read 2611.13907 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper