AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers

Wenhan Ma, Hailin Zhang, Liang Zhao, Yifan Song, Yudong Wang, Fuli Luo, Zhifang Sui

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
add Rollout Routing Replay (R3)MoE inference-training KL · Qwen3-30B-A3B training vs inferenceKL divergence: 1.5 × 10^-3 to 7.5 × 10^-4Improved
add R3 to GRPOGRPO · Qwen3-30B-A3B-SFT mini_step=8average score: 48.84 -> 68.05Improved
replace GSPO with GRPO+R3GSPO · Qwen3-30B-A3B-SFT mini_step=8average score: 66.76 -> 68.05Improved
add R3 to GSPOGSPO · Qwen3-30B-A3B-SFT mini_step=8average score: 66.76 -> 69.00Improved
replace TIS with R3GRPO+TIS · Qwen3-30B-A3B-SFT mini_step=1average score: 66.24 -> 71.83Improved
replace TIS with R3GRPO+TIS · Qwen3-30B-A3B-Base mini_step=1average score: 69.22 -> 70.73Improved
add TIS on top of R3 (ablation)GRPO+R3 · Qwen3-30B-A3B-SFT mini_step=1average score: 71.83 -> 70.14Hurt
add TIS on top of R3 (ablation)GRPO+R3 · Qwen3-30B-A3B-Base mini_step=1average score: 70.73 -> 69.17Hurt
add R3 to GRPOGRPO · SWE-bench VerifiedPass@1: 31.80 -> 38.60Improved
add R3 to GRPOGRPO · Qwen3-30B-A3B-ReasoningSFT mini_step=4 (vs GSPO)average score: 74.62 -> 77.00Improved
add R3 to GRPOGRPO · Qwen3-30B-A3B-ReasoningSFT mini_step=1average score: 75.30 -> 76.79Improved
add R3 to GRPOGRPO · Qwen3-30B-A3B-Base mini_step=1average score: 61.69 -> 70.73Improved

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