TurnWise: turn-level GRPO improves when-to-answer in streaming video RL
_akhaliq · x · 2026-11-06
TurnWise is a turn-level GRPO framework giving each decision in a streaming conversation its own credit—rewarding turns that add information and penalizing repeats—instead of sharing one signal across all decisions. It outperforms prior RL methods on three streaming video benchmarks.
Related event: TurnWise: Turn-Level Credit Assignment for Streaming Video RL(2 posts)→
More from Research
- PhAI Labs unveils Discovery Foundation Models to bring AI into scientific discovery — rohanpaul_ai · 2026-11-06
- MetricTrail: Pose Estimation for Long Videos Without Long-Sequence Training — AntonObukhov1 · 2026-11-06
- Oxford Professor Robin Evans, Causal Inference Scholar, Dies After Illness — murat_kocaoglu_ · 2026-11-06
- Jev 公开 72 万条决策数据,可自训轻量级路由/决策模型 — MaziyarPanahi · 2026-11-06
- RespSurvBench: top model passes just 25.5% of respiratory surveillance evals — kenbwork · 2026-11-06
- ARC-AGI-3 demands extreme sample efficiency, unlike the brute-force Atari RL era — Ryan__Rhys · 2026-11-06