2026-11-05 · America/Los_Angeles · 论文 · #2

Turns Matter: Advancing Streaming Video RL via Turn-Level Credit

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 trajectory-level advantage across turns. Such shared credit can reinforce redundant replies after earlier successes and discourage useful follow-ups after earlier failures. We propose TurnWise, a turn-level GRPO framework that aligns learning with individual streaming decisions. TurnWise addresses the mismatch between…

热度 53.5 / 100；排名与评分保留该期记录。