When Does Muon Help Agentic Reinforcement Learning?

[Submitted on 17 Jul 2026 (v1), last revised 20 Jul 2026 (this version, v2)]

View PDF HTML (experimental)

Abstract:Muon is competitive with AdamW in large-scale pre-training, but its value for reinforcement-learning (RL) post-training remains unclear. We study vanilla Muon in sparse-reward agentic RL through matched single-seed comparisons with AdamW on ALFWorld using Qwen2.5-0.5B-Instruct. Under Group-in-Group Policy Optimization (GiGPO), applying Muon only to hidden weight matrices raises final-window validation success from 0.290 to 0.546 (+88%); high-rate AdamW controls retain no post-update success. The effect depends on the advantage estimator and learning rate. At 3e-5, Muon improves GRPO from 0.161 to 0.268, whereas GraphGPO’s late-window gap narrows near saturation. At 1e-5, GraphGPO Muon reaches 0.901, raises normalized validation AUC from 0.399 to 0.556, and reaches 0.5 and 0.75 success 30 and 60 updates earlier, respectively. These exploratory results show that Muon can benefit agentic RL and motivate studying the policy optimizer, advantage estimator, and learning rate jointly.

Submission history

From: Kai Ruan [view email]
[v1] Fri, 17 Jul 2026 17:49:05 UTC (737 KB)
[v2] Mon, 20 Jul 2026 17:21:45 UTC (815 KB)