/Tech3h ago

Muon Optimizer Beats Adam on Sequenced LLM Code Training Tasks

116161.1K

Original post

here's the Muon-OGD paper's claimed numbers for sequenced task domain training on Adam vs their Muon variant btw it's uh, not at all a small gap.

kalomaze@kalomaze

>claim: these results are "general and fundamental" to the architecture >they exclusively tested with Adam hmm, interesting work, but... you really can't study this purely in the diagonal preconditioning regime.

12:26 AM · Jun 25, 2026 · 561 Views

Sentiment

Sentiment building, check back later.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

ARXIV.ORGVia

#1213

Posts from X

Most Activity

VIEWS507BOOKMARKS4LIKES7RETWEETS1

kalomaze@kalomaze

from: https://arxiv.org/abs/2605.08949

kalomaze@kalomaze

here's the Muon-OGD paper's claimed numbers for sequenced task domain training on Adam vs their Muon variant btw it's uh, not at all a small gap.

3h50774