AI Paper Atlas

Research map / Optimization / Optimizers

Towards Efficient Optimizer Design for LLM via Structured Fisher Approximation with a Low-Rank Extension

Wenbo Gong, Meyer Scetbon, Chao Ma, Edward Meeds

ICLR 2026 · Optimization › Optimizers · approach: Memory-efficient LLM optimizers · 7 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add optimal low-rank compensation (ablation)Alice · C4 eval ppl, LLaMA-130Meval perplexity: 25.11 -> 21.95Improved
add subspace switching on top of tracking (ablation)Alice · C4 eval ppl, LLaMA-130Meval perplexity: 27.35 -> 25.11Improved
add low-rank tracking alone (no switching or compensation) (ablation)Alice · C4 eval ppl, LLaMA-130Meval perplexity: 26.96 -> 27.35Hurt
drop tracking (Alice-0) in the full Alice (ablation)Alice · C4 eval ppl, LLaMA-130Meval perplexity: 21.95 vs 21.99No clear effect
use orthogonal-complement switching instead of Gaussian or Gaussian-mix bases (ablation)Alice · C4 eval ppleval perplexityImproved
use optimal compensation instead of Fira's heuristic compensation (ablation)Alice · C4 eval ppleval perplexityImproved
use Alice instead of AdamAlice · LLaMA-1.3B C4 pretrainingeval perplexity: 16.44 -> 13.52Improved
use Alice instead of FiraAlice · LLaMA-130M C4, last layer not Adameval perplexity: 25.21 -> 21.95Improved
use RACS (SSGD) instead of Apollo-svdRACS · LLaMA-1.3B C4eval perplexity: 14.10 -> 13.52Improved
train 1B LLaMA with Alice instead of 7B with ApolloAlice · C4 eval ppl at 150K stepseval perplexity: 13.02 -> 12.97Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy