AMO adapts Newton-Schulz iteration counts per operator type using early-training geometry measurements to improve Muon optimizer performance over uniform schedules in LLM pre-training.
Root mean square layer normalization
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
AMO: Adaptive Muon Orthogonalization
AMO adapts Newton-Schulz iteration counts per operator type using early-training geometry measurements to improve Muon optimizer performance over uniform schedules in LLM pre-training.