AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Experience Augmented Policy Optimization for LLM Reasoning

Jinda Lu, Kexin Huang, Junkang Wu, Shuo Yang, Jinghan Li, Chiyu Ma, Shaohang Wei, Xiang Wang, Guoyin Wang, Jingren Zhou

ICML 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards · 0 citations (OpenAlex)

Read the paper

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 smoothed response-level importance sampling to EAPO (ablation)EAPO · Math average (Qwen-2.5-Math-7B)Pass@1 (%): 42.76 -> 43.34Improved
add smoothed importance sampling to EAPO (ablation)EAPO · GPQA (Qwen-2.5-Math-7B)accuracy (%): 41.98 -> 43.18Improved
add smoothed importance sampling to EAPO (ablation)EAPO · Math average (Qwen-2.5-Math-7B)Pass@16 (%): 62.37 -> 62.41No clear effect
use token-level experience instead of trajectory-level experience (positive samples only) (ablation)EAPO · AIME'24Mean@32Improved
include negative experience-augmented samples in addition to positive-only filtering (ablation)EAPO · AIME'24Mean@32Hurt
raise resampling threshold tau to 0.8 or above (ablation)EAPO · AIME'24Mean@32Hurt
vary resampling threshold tau within 0.3 to 0.5 (ablation)EAPO · AIME'24Mean@32: 0.381 at tau=0.5No clear effect
use EAPO experience injection instead of plain DAPODAPO · Math average (Qwen-3-8B-Base)Pass@1 (%): 45.31 -> 49.71Improved
use EAPO instead of DAPODAPO · Math average (Qwen-2.5-Math-7B)Pass@1 (%): 40.56 -> 43.34Improved
use EAPO instead of DAPODAPO · GPQA and MMLU-Pro (Qwen-2.5-Math-7B)average accuracy (%): 41.68 -> 45.38Improved
replace token-level resampling with trajectory replay of successful trajectoriesEAPO · Math average (Qwen-2.5-Math-7B)Pass@1 (%): 43.34 -> 42.08Hurt
add on-policy distillation from RL-optimized policy (OPD) instead of EAPODAPO · Math average (Qwen-2.5-Math-7B)Pass@1 (%): 40.56 -> 40.64No 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