Video models improve accuracy by keeping first answers confident
Confident First, Reflective Second: Reinforcement Learning of Confidence-Preserving Reflection for Video Understanding
Computer Vision and Pattern RecognitionMultimedia
Summary
Video understanding models often look at a few frames and give an answer but might miss important moments. The authors found that if models are trained to rethink their answers, the first answer can actually become less reliable. They propose a new training method called CoFire that makes sure the first guess stays trustworthy and that reflection only improves the answer. This approach leads to better overall accuracy when understanding videos.
What this means in practice
- •For video app developers: Build video apps that analyze clips more accurately by keeping early predictions reliable before refining them.
- •For media content moderation teams: Flag videos more effectively by ensuring initial model assessments are solid before applying further review steps.
Authors
Zihao Lin, Jingyuan Liu, Yucheng Zhang, Yuguang Yao, Lifu Huang
Abstract
Video models answer from sampled frames and may miss the decisive moment. A common remedy is to commit to an answer, re-examine the video, and confirm or revise it, a loop now mainly learned with reinforcement learning (RL). We find that RL erodes the initial answer, a phenomenon we term confidence erosion, where confidence means the label-verified reliability of the initial answer instead of self-reported certainty: because errors can be repaired later, a correct initial answer earns less reward, and the initial answer stalls even as the final answer's gain over it rises. We propose the principle confident first, reflective second: the initial answer must stay reliable on its own, above a floor set by a reference policy, and reflection may only add to it. Our method, CoFire, implements this principle with one mechanism for each erosion channel: reward incentives, parameter drift, and gradient interference. CoFire reaches 67.0% mean accuracy on Video-MME, MLVU, and LongVideoBench, 1.7 points above the strongest compared 7B method, and improves initial-answer and final accuracy over outcome-reward RL by 4.1 and 3.7 points on average: a stronger initial answer carries through to a stronger final one.