TurnWise：基于回合级信用的流式视频强化学习

AI 摘要

该论文提出 TurnWise，一种面向流式视频理解的回合级 GRPO 强化学习框架。现有方法在多个回合间共享轨迹级优势，容易在早期成功后强化冗余回复、在早期失败后抑制有用的后续响应。TurnWise 通过静默补全前缀评分实现回合级信用分配，并用回合均值策略优化让回复与等待决策等权。实验表明该方法优于此前的流式视频理解强化学习方法。

正文节选

Turns Matter: Advancing Streaming Video RL via Turn-Level Credit Abstract TurnWise improves streaming video understanding by assigning turn-level credit and balancing decision weights in group relative policy optimization. Streaming video understanding requires models to repeatedly decide whether to respond or continue observing. Recently, Group Relative Policy Optimization (GRPO) has been applied to this setting, with existing approaches emphasizing reward design while retaining a single trajec

发布时间：—

抓取时间：2026-11-06 00:18

来源机构：Hugging Face