Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-06-23 · source-backed
Hübler, Lion, and Orvieto analyze the matrix-aware Muon/Muown optimizer family and show Muown's edge comes from an implicit angular step-size decay, giving a mechanistic reason these optimizers beat standard methods with practical implications for learning-rate scheduling. (arXiv) Niche, but if you're experimenting with Muon-style optimizers in pretraining, this explains the "why" instead of leaving it as folklore. Also worth a look this week: scaling linear mode connectivity and weight-merging to billion-parameter Transformers (arXiv), useful if you're merging fine-tunes without retraining.
Each link below shares sources, entities, or timing with this story.
Hugging Face released transformers / Shared entity: Transformers / What happened next
Linked by a graph relationship (Hugging Face released transformers); both cover Transformers; picks up the Transformers thread on 2026-07-09.
Shared entity: Muon / Same source domain / Earlier coverage / Tension
Both cover Muon; reported by the same outlet (arxiv.org); earlier Muon coverage from 2026-03-19.
Shared entity: Muon / Same source domain / What happened next
Both cover Muon; reported by the same outlet (arxiv.org); picks up the Muon thread on 2026-08-06.
Shared entity: Transformers / Same source domain / What happened next
Both cover Transformers; reported by the same outlet (arxiv.org); picks up the Transformers thread on 2026-08-03.
Both cover Transformers; reported by the same outlet (arxiv.org); picks up the Transformers thread on 2026-07-20.
Hugging Face released transformers / Same source domain
Linked by a graph relationship (Hugging Face released transformers); reported by the same outlet (arxiv.org).
Linked by a graph relationship (Hugging Face released transformers); reported by the same outlet (arxiv.org).
Linked by a graph relationship (Hugging Face released transformers); reported by the same outlet (arxiv.org).