AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Implicit Turn-Wise Policy Optimization for Proactive User-LLM Interaction

Haoyu Peter Wang, Yuxin Chen, Liang Luo, Buyun Zhang, Ellie Dingqiao Wen, Pan Li

ICML 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Aggregate token rewards into turn-wise rewards with softmax normalization (Norm-ITPO) instead of sparse trajectory rewardRLOO, Qwen2.5-3B-Instruct · Math Tutoringaccuracy: 29.06 -> 32.50Improved
Add Norm-ITPO turn rewards over trajectory-level rewardRLOO, Qwen2.5-3B-Instruct · Medical Recommendationdiagnosis score: 65.12 -> 69.24Improved
Add Norm-ITPO turn rewards over trajectory-level rewardGRPO, Qwen2.5-3B-Instruct · Medical Recommendationdiagnosis score: 66.41 -> 71.74Improved
Add Norm-ITPO turn rewards to PPO value modelPPO, Qwen2.5-3B-Instruct · Doc WritingBLEU: 40.57 -> 45.16Improved
Add reward normalization (Norm-ITPO vs ITPO) (ablation)PPO with value model · Medical Recommendationoverall score: 54.87 -> 59.44Improved
Add reward normalization (Norm-ITPO vs ITPO) (ablation)RLOO · Medical Recommendationdiagnosis score: 68.43 -> 69.24Improved
Use token-level implicit rewards (PRIME) instead of turn-wise (ablation)RLOO · Doc WritingBLEU: 44.59 -> 40.95Hurt
Use token-level implicit rewards (PRIME) instead of turn-wise (ablation)PPO value model · Math Tutoringaccuracy: 26.81 -> 22.63Hurt
Train an extra value model (PPO) instead of group-based advantagesQwen2.5-3B-Instruct · Math Tutoring, Doc Writing, Medical Recommendationaccuracy / BLEU / diagnosisNo clear effect
Uniform random decomposition of outcome rewardRLOO · Math Tutoring, Doc Writing, Medical Recommendationaccuracy / BLEU / diagnosisNo clear effect
Use LLM-as-Judge turn weightsRLOO · Doc WritingBLEU: 37.35 -> 42.34Improved
Apply Norm-ITPO to other policy backbonesRLOO, Qwen3-4B and Qwen2.5-7B · Doc WritingscoreImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy