AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Balancing the Experts: Unlocking LoRA-MoE for GRPO via Mechanism-Aware Rewards

Changlian Ma, Zizheng Huang, Xiangyu Zeng, Yi Wang, Cheng Liang, Kun Tian, Xinhai Zhao, Limin Wang

ICLR 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants

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 balancing reward RB to RO-GRPO (Smooth) (ablation)RO-GRPO (Smooth) · GSM8K (E=2)accuracy (%): 90.75 -> 91.51Improved
add confidence reward RH to RO-GRPO (Smooth) (ablation)RO-GRPO (Smooth) · GSM8K (E=2)accuracy (%): 89.92 -> 91.51Improved
add balancing reward RB to RO-GRPO (Relative) (ablation)RO-GRPO (Relative) · GSM8K (E=2)accuracy (%): 89.01 -> 90.22Improved
add confidence reward RH to RO-GRPO (Relative) (ablation)RO-GRPO (Relative) · GSM8K (E=2)accuracy (%): 90.14 -> 90.22No clear effect
use correct rather than shuffled routing rewards (ablation)RO-GRPO (Smooth) · GSM8K (E=2)accuracy (%): 89.23 -> 91.51Improved
use correct rather than shuffled routing rewards (ablation)RO-GRPO (Relative) · GSM8K (E=2)accuracy (%): 89.28 -> 90.22Improved
add routing reward (Smooth) to GRPO LoRA-MoEGRPO LoRA-MoE 4x8 · Geometry3kaccuracy (%): 28.95 -> 40.10Improved
add routing reward (Relative) to GRPO LoRA-MoEGRPO LoRA-MoE 4x8 · Geometry3kaccuracy (%): 28.95 -> 40.27Improved
add routing reward to reduce repetitive-loop failuresGRPO LoRA-MoE 4x8 · Geometry3krepetitive-loop rate: 7.5% -> 0.17% (Smooth), 0% (Relative)Improved
replace auxiliary routing loss with routing reward (Smooth)LoRA-MoE with GRPO · GSM8K (E=2)accuracy (%): 86.73 -> 91.51Improved
replace auxiliary routing loss with routing rewardLoRA-MoE with GRPO · GSM8K (E=2)load-balancing MSE: 0.036 -> 0.016Improved
add auxiliary routing loss to GRPO LoRA-MoEGRPO LoRA-MoE 2x8 · GSM8Kaccuracy (%): 89.39 -> 86.73Hurt

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