Research map / Reinforcement learning / RL from feedback
EAPO: Enhancing Policy Optimization with On-Demand Expert Assistance
Siyao Song, Cong Ma, Zhihao Cheng, Shiye Lei, Minghao Li, Ying Zeng, Huaixiao Tou, Kai Jia
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add on-demand expert assistance in RL (EAPO) vs self-exploratory RL | R1-Distill-7B RL · Math average | Pass@32: 59.16 -> 64.07 | Improved |
| add expert assistance in RL on Qwen3-8B | Qwen3-8B RL · Math average | Pass@32: 45.44 -> 49.98 | Improved |
| use parallel instead of sequential expert queries (ablation) | EAPO · Math average | Pass@32: 61.79 -> 64.07 | Improved |
| use 32B expert instead of 14B expert (ablation) | EAPO · Math average | Pass@32: 61.55 -> 64.07 | Improved |
| use heterogeneous Llama-8B experts instead of homogeneous R1-distill-7B experts (ablation) | EAPO · Math average | pass@32: 58.85 -> 60.66 | Improved |
| add homogeneous expert pool to self-exploratory RL (ablation) | EAPO · Math average | pass@32: 59.16 -> 58.85 | No clear effect |
| use annealed schedule s^-1 instead of budget-matched fixed rate 0.02 (ablation) | EAPO · Math average | Pass@32: 59.65 -> 64.07 | Improved |
| use faster decay s^-2 instead of s^-1 (ablation) | EAPO · Math average | Pass@32: 64.07 -> 59.76 | Hurt |
| use informative random-sampled expert guidance instead of error-sampled guidance (ablation) | Expert-Assisted Workflow · Math average | pass@32: 42.82 -> 49.39 | Improved |
| use real expert guidance instead of meaningless guidance (ablation) | Expert-Assisted Workflow · Math average | pass@32: 42.92 -> 49.39 | Improved |
| self collaboration with 7B copies | R1-7B · Math average | accuracy: 42.53 -> 41.83 | No clear effect |
| use stronger expert (DeepSeek_V3.2) over Qwen3-235B-A22B in workflow (ablation) | Expert-Assisted Workflow · Math average | accuracy: 52.61 -> 54.55 | Improved |
Later papers that compared with this method
1 later accepted paper reports EAPO as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability (NeurIPS 2026 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DAPO : Improving Multi-Step Reasoning Abilities of Large Language Models with Direct Advantage-Based Policy Optimization (NeurIPS 2025 · similarity 0.91)
- EEPO: Exploration-Enhanced Policy Optimization via Sample-Then-Forget (ICLR 2026 · similarity 0.91)
- Agentic Reinforced Policy Optimization (ICLR 2026 · similarity 0.89)
- Selective Expert Guidance for Effective and Diverse Exploration in Reinforcement Learning of LLMs (ICLR 2026 · similarity 0.88)
- Accelerating RL for LLM Reasoning with Optimal Advantage Regression (NeurIPS 2025 · similarity 0.88)
- One-Way Policy Optimization for Self-Evolving LLMs (ICML 2026 · similarity 0.88)
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.
