TurnWise RL Framework Cuts Video Agent Repetition by 61.5%

A 10-institution research team led by Fudan University and JD.com has released TurnWise, a reinforcement learning framework that assigns credit at the turn level rather than per-trajectory, achieving a 61.5 percent reduction in streaming video agent repetition. The method targets a longstanding credit-assignment pathology that has prevented streaming video RL from matching the performance of reasoning-heavy text RL systems. Code, a tuned 4B checkpoint, and the mmduet2 training dataset of 3,983 streaming instances are now available on Hugging Face.

The Credit-Assignment Pathology

Streaming video agents face a unique decision structure: at every turn, the model must choose whether to answer or stay silent. Standard Group Relative Policy Optimization scores the entire trajectory and broadcasts a single advantage signal back to every turn. According to the research team, this creates a systematic problem. After a successful early response, a redundant follow-up reply may inherit a positive trajectory advantage even when the reward function penalizes repetition. Conversely, a useful follow-up after earlier failures may receive negative advantage because the overall trajectory remains below the group average.

In a cohort of 11,116 training trajectories, the researchers found that turn-level credit flipped 432 of 722 repetitive replies, or 59.83 percent, from positive to negative learning signals. These were replies that trajectory-level GRPO had been rewarding despite their redundancy.

How TurnWise Works

TurnWise rewrites the credit step by rescoring every prefix after replacing all later actions with silence, then using the differences to compute group-relative remaining-return advantages per decision. The framework pairs this with turn-mean loss aggregation, which averages token losses inside a turn before averaging across turns. This ensures a long response and a short NO REPLY carry equal weight in training. The method requires neither a learned critic nor additional rollouts, according to the paper.

Benchmark Performance

On Qwen2.5-VL-3B, the trained TurnWise policy reaches 43.56 PAUC on ProactiveVQA, 21.59 average Timing F1 on OVO-Timing, and 49.43 average accuracy on StreamingBench. These results exceed the strongest non-TurnWise baseline by 3.96, 3.87, and 0.69 points respectively. On the larger Qwen3-VL-4B backbone, the gains over the strongest baseline grow to 6.26, 1.92, and 2.38 points on the same three benchmarks. The combination of turn-level credit and turn-mean aggregation dropped the repetition rate on the 3B backbone from 15.53 to 5.98.

Release Date and Availability

Officially released on November 5, 2026, TurnWise arrives as streaming video understanding becomes a priority research area. The full codebase, the tuned Qwen3-VL-4B checkpoint, and the mmduet2 training set are published alongside the preprint on Hugging Face. The research team reports no wall-clock cost increases against prior methods including MMDuet2, ThinkStream, and StreamPro, though per-category StreamingBench numbers for the 4B backbone are not broken out in the results section.

What This Means

TurnWise demonstrates that fine-grained credit assignment can address a core failure mode in streaming video RL without introducing computational overhead or requiring learned critics. By flipping nearly 60 percent of repetitive replies from positive to negative learning signals, the framework brings streaming video agent training closer to the credit-assignment precision that has driven progress in text-based reasoning systems. The immediate availability of code and trained checkpoints lowers the barrier for research teams working on multimodal agents and real-time video understanding.