AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Master Skill Learning with Policy-Grounded Synergy of LLM-based Reward Shaping and Exploring

Yanbin Chang, Junfan Lin, Jie Jiang, Runhao Zeng, Changxin Huang, Jianqiang Li

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning

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
add goal-oriented reward Rg (ablation)PoRSE · TwoCatchMTS: 0.190 -> 0.349Improved
add LLM-designed exploration bonus Re (ablation)PoRSE · BlockStackMTS: 0.393 -> 0.753Improved
add dynamic policy fusion inheritance (ablation)PoRSE · TwoCatchMTS: 0.017 -> 0.349Improved
add dynamic policy fusion inheritance (ablation)PoRSE · FrankaCabinetMTS: 0.671 -> 0.957Improved
adapt reward ratio beta (vs fixed) (ablation)PoRSE · PushBlockMTS: 0.243 -> 0.318Improved
use task-aware LLM affordance mapping instead of SimHash (ablation)PoRSE · 8 Bi-DexHands tasksMTS: SwingCup 0.454 -> 0.995; DoorOpenInward 0.000 -> 0.250Improved
use adaptive refined bonus instead of fixed LLMCount bonus (ablation)PoRSE · TwoCatchMTS: 0.000 -> 0.349Improved
use fused Rg+Re versus single reward (ablation)PoRSE · DoorOpenInward / TwoCatchsuccess rate: 0% -> 97% (DoorOpenInward), 25% (TwoCatch)Improved
replace LLM-generated AFS with random AFS (ablation)PoRSE · BlockStackMTS: 0.753 -> 0.680Hurt
replace DeepSeek-V3 with GPT-4o-mini (ablation)PoRSE · BlockStackMTS: 0.753 -> 0.618Hurt
add prompt-guided task-irrelevant AFS dimensions (ablation)PoRSE · FrankaCabinetMTS: 0.957 -> 0.951No clear effect
increase bin count from 1000 to 3000 (ablation)PoRSE · PushBlockMTS: 0.378 -> 0.321No clear effect

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