[HUGGINGFACE]score: 0.71
Pion Optimizer Fixes Muon's Spectral Whitening Failures in VLA and RLVR Training
May 18, 2026
Muon's uniform spectral whitening amplifies noisy gradient directions in low-rank action modules (VLA) and destabilizes per-head specialization in RLVR; Pion replaces it with high-pass spectral filtering as a drop-in fix for both regimes.
paper