Research map / Reinforcement learning / Policy gradient and actor-critic
GEPO: Group Expectation Policy Optimization for Stable Heterogeneous Reinforcement Learning
Han Zhang, RuibinZheng, ZEXUAN YI, Zhuo Zhang, Hanyang Peng, Hui Wang, Jiayin Qi, Binxing Fang, Ruifeng Xu, Yue Yu
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 |
|---|---|---|---|
| use group-level importance weights (GEPO) instead of token-level (GRPO) | GRPO · Online RL, Qwen3-8B, 4k, average | avg Pass@1: 61.8 -> 65.0 | Improved |
| use group-level importance weights (GEPO) instead of token-level (GRPO) | GRPO · Online RL, Qwen3-1.7B, 4k, average | avg Pass@1: 49.9 -> 51.1 | Improved |
| use group-level instead of sequence-level importance weights | GSPO · Hetero RL max delay 64, Qwen3-8B, last checkpoint | avg Pass@1: 46.4 -> 60.8 | Improved |
| use group-level instead of sequence-level importance weights | GSPO · Hetero RL max delay 64, Qwen3-8B, best checkpoint | avg Pass@1: 58.4 -> 62.6 | Improved |
| use group-level instead of token-level importance weights | GRPO · Hetero RL max delay 64, Qwen3-8B, last checkpoint | avg Pass@1: 55.8 -> 60.8 | Improved |
| use sequence-level instead of token-level weights | GRPO · Hetero RL max delay 64, last checkpoint | avg Pass@1: 55.8 -> 46.4 | Hurt |
| use sequence-level instead of token-level weights | GRPO · Online RL, Qwen3-8B, 4k, average | avg Pass@1: 61.8 -> 60.5 | No clear effect |
| use BNPO or Dr.GRPO instead of GRPO | GRPO · Hetero RL max delay 64, Qwen3-8B | avg Pass@1: 56.6 -> 55.9 (BNPO), 56.0 (Dr.GRPO) | No clear effect |
| increase network latency (max delay 8 to 64 steps) (ablation) | GEPO · Hetero RL training dynamics | KL divergence / eval accuracy | Hurt |
| higher latency raises KL divergence and importance weight variance | importance sampling · Hetero RL, correlation analysis | correlation: r = 0.96 (IW variance vs KL) | Hurt |
| use group-expectation weight instead of 1/q | importance weights · Bias analysis | mean bias: 0.075 -> 0.072 | No clear effect |
| set KL coefficient to 0.005 and group size 8 (ablation) | GEPO · Hetero RL hyperparameter sweep, Qwen3-1.7B, 2k | avg best/last | Improved |
Later papers that compared with this method
1 later accepted paper reports GEPO as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Phase-Aware Mixture of Experts for Agentic Reinforcement Learning (ICML 2026 · better on 3 of 3 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
- Group-in-Group Policy Optimization for LLM Agent Training (NeurIPS 2025 · similarity 0.87)
- Asynchronous Federated Reinforcement Learning with Policy Gradient Updates: Algorithm Design and Convergence Analysis (ICLR 2025 · similarity 0.87)
- Heterogeneous Agent Collaborative Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (NeurIPS 2026 · similarity 0.87)
- Momentum for the Win: Collaborative Federated Reinforcement Learning across Heterogeneous Environments (ICML 2024 · similarity 0.87)
- SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores (ICLR 2024 · similarity 0.87)
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.
