Muon holds its own against AdamW when pre-training at scale. Whether that advantage carries into reinforcement learning post-training is a separate question, and this study tests it in the specific setting of sparse-reward agentic RL on ALFWorld with a small Qwen model.

The interesting result is conditional rather than general. Applying Muon only to hidden weight matrices lifted final-window validation success substantially, which is a finding about where the optimiser helps rather than whether it helps. Different parameter groups apparently want different treatment.

The methodology deserves a note too: matched single-seed comparisons are honest about their own limits. One seed cannot separate a real effect from variance, and reporting it plainly is better than the alternative common in this literature, which is to run several and present the flattering one.