Turn-level credit lets a streaming video model learn each reply-or-wait decision separately, instead of sharing one reward across the whole stream.
Editorial summary · not the author abstract. Read the linked paper for the complete abstract and claims.Figure 1: a shared trajectory-level advantage can reinforce redundant replies; TurnWise assigns credit to each turn. Original figure · paper authors ↗
Release evidence
Public repository has RL training code; a 4B checkpoint and frame-extracted training data are on Hugging Face. The checkpoint card does not state its base model. Not rerun.