Spectral deflation framework improves Muon: consistent validation loss gains in GPT-2 pretraining
hankyang94 · x · 2026-09-28
New work by Haoran Sun and Shucheng Kang introduces Spectral Deflation, a general framework for factorization-free matrix filtering in Muon and semidefinite programming.
- Shared bottleneck: fixed-depth polynomial matrix filters — used in Muon-based LLM pretraining and ADMM-based SDP — suffer from a few dominant spectral components compressing the rest of the spectrum after normalization, hurting filter accuracy.
- Method: deflate the dominant components, generalized into a spectral deflation framework applied to both settings.
- Results: consistently improves validation loss in GPT-2 pretraining with Muon, and substantially reduces the KKT residual of factorization-free ADMM at comparable projection time.
More from Infra
- Sarvam AI Launches Token Factory: India-Hosted Inference for Open Models — itsOmSarraf_ · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Stratechery's Fictional Obituary: Imagining Jensen Huang's 2026 Exit From Nvidia — Gregory_C_Allen · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16