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

**Authors:** Shuai Dong, Kanghui Tian, Linghan Chen, Wenqi Shao, Hongrui Wu, Haowen Hou, Yongfu Zhu, Xuehui Wang, Ziyue Wang, Wenpu Liu, Weichu Xie, Yuqi Xu, Congcong Wang, Kaiwen Tuo, Caoyuan Ma, Taiqiang Wu, Ji Zhao, Wenzheng Chang, Dianyi Wang, Siyuan Wang, Zhongyu Wei, Yu Qiao, Jiaqi Wang  
**Lab:** JD.com · **Published:** 4 Nov 2026 · **Upvotes:** 53 · **Level:** Intermediate  
**Links:** [arXiv](https://arxiv.org/abs/2611.06728) · [PDF](https://arxiv.org/pdf/2611.06728) · [Code](https://github.com/XD111ds/turn-wise-rl) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.06728.html)

## Curator's notes

**TL;DR:** TurnWise adds turn-level credit assignment to streaming video RL, improving decision quality without extra critics.

- **What's new:** Introduces turn-level GRPO with silence-completed prefix scoring and turn-mean policy optimization, unlike prior single trajectory-level advantage methods.
- **Key result:** Extensive experiments show TurnWise outperforms prior RL approaches for streaming video understanding.
- **Why it matters:** It yields more informative, less redundant streaming responses, benefiting real-time video applications.

## Abstract

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 trajectory-level feedback and turn-level interaction, promoting informative responses and reducing redundancy. Specifically, turn-level credit assignment uses silence-completed prefix scoring to construct group-relative advantages from each decision's remaining return, separating its learning signal from earlier rewards. Turn-mean policy optimization averages token losses within each turn and then across turns, giving response and waiting decisions equal weight regardless of output length. These complementary designs improve streaming policy learning without a learned critic or additional rollouts. Extensive experiments show our method outperforms prior RL approaches for streaming video understanding.

## What you could build

1. Implement a live video captioning system using TurnWise for adaptive response timing
2. Create a streaming video QA bot that decides when to answer versus wait using turn-level credit
3. Develop a real‑time video summarizer that leverages TurnWise to balance silence and output

## Quiz yourself

How does TurnWise assign credit to each decision unlike previous methods?

**Answer:** It uses silence-completed prefix scoring to compute group‑relative advantages from the remaining return of each turn, separating the learning signal from earlier rewards.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.