DynMuon dynamically schedules the spectral exponent p in Muon-style updates according to curvature, noise, and training stage, yielding lower validation loss with 10-26% fewer steps than fixed Muon.
modded-nanogpt: Speedrunning the nanogpt baseline
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
fields
cs.LG 2years
2026 2verdicts
UNVERDICTED 2representative citing papers
The same Transformer architecture follows different spectral scaling laws under different optimizers, with Muon achieving linear hard-rank scaling on tail representations while AdamW shows weak scaling, even when perplexity is matched.
citing papers explorer
-
DynMuon: A Dynamic Spectral Shaping View of Muon
DynMuon dynamically schedules the spectral exponent p in Muon-style updates according to curvature, noise, and training stage, yielding lower validation loss with 10-26% fewer steps than fixed Muon.
-
Same Architecture, Different Capacity: Optimizer-Induced Spectral Scaling Laws
The same Transformer architecture follows different spectral scaling laws under different optimizers, with Muon achieving linear hard-rank scaling on tail representations while AdamW shows weak scaling, even when perplexity is matched.