TurnWise Targets Streaming Video Decisions

Researchers propose TurnWise, a turn-level reinforcement learning framework for streaming video understanding, in a paper published on November 4 and submitted to Hugging Face on November 5. The method addresses a repeated decision in streaming video: whether a model should respond or continue observing. The paper says Group Relative Policy Optimization (GRPO) has been used for this setting, but existing approaches focus on reward design while assigning one shared advantage across an entire trajectory of turns. That shared credit can reinforce redundant replies after an earlier success and discourage useful follow-ups after an earlier failure.

TurnWise assigns credit to each individual decision, aligning learning signals with the turn-by-turn interaction. Its turn-level credit assignment uses silence-completed prefix scoring to calculate group-relative advantages from the remaining return after each decision, separating that signal from rewards earned earlier. The framework also uses turn-mean policy optimization: it averages token losses within each turn, then averages across turns, so response and waiting decisions receive equal weight regardless of how many tokens they produce.

The authors say the two designs improve streaming policy learning without a learned critic (a model that estimates future reward) or additional rollouts (extra sampled interaction sequences). They report that extensive experiments found TurnWise outperformed prior reinforcement learning approaches for streaming video understanding. The paper lists 22 authors and appeared as the #2 Paper of the day on Hugging Face, with 89 upvotes shown on the page.

Researchers propose TurnWise, a turn-level reinforcement learning framework for streaming video understanding, in a paper published on November 4 and submitted to Hugging Face on November 5. The method addresses a repeated decision in streaming video: whether a model should respond or continue observing. The paper says Group Relative Policy Optimization (GRPO) has been used for this setting, but existing approaches focus on reward design while assigning one shared advantage across an entire trajectory of turns. That shared credit can reinforce redundant replies after an earlier success and discourage useful follow-ups after an earlier failure.

TurnWise assigns credit to each individual decision, aligning learning signals with the turn-by-turn interaction. Its turn-level credit assignment uses silence-completed prefix scoring to calculate group-relative advantages from the remaining return after each decision, separating that signal from rewards earned earlier. The framework also uses turn-mean policy optimization: it averages token losses within each turn, then averages across turns, so response and waiting decisions receive equal weight regardless of how many tokens they produce.

The authors say the two designs improve streaming policy learning without a learned critic (a model that estimates future reward) or additional rollouts (extra sampled interaction sequences). They report that extensive experiments found TurnWise outperformed prior reinforcement learning approaches for streaming video understanding. The paper lists 22 authors and appeared as the #2 Paper of the day on Hugging Face, with 89 upvotes shown on the page.