AI Paper Atlas

Research map / Optimization / Optimizers

FOAM: Blocked State Folding for Memory-Efficient LLM Training

Ziqing Wen, Jiahuan Wang, Ping Luo, Dongsheng Li, Tao Sun

ICML 2026 · Optimization › Optimizers · approach: Memory-efficient LLM optimizers · 0 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 residual correction term R_t to blocked-average update (ablation)FOAM · C4 pre-training LLaMA-60M/130Mvalidation perplexityImproved
use FOAM-3 (level 3) instead of FOAM-2 (ablation)FOAM · C4 LLaMA-1.3Bperplexity: 13.13 -> 13.19; memory 4.45G -> 3.97GMixed
vary level l (higher means less memory) (ablation)FOAM · C4 pre-trainingvalidation perplexityNo clear effect
replace Full-Adam with FOAM-2Full-Adam · C4 LLaMA-1.3Bperplexity: 14.51 -> 13.13Improved
replace APOLLO-1/4 with FOAM-2APOLLO · C4 LLaMA-350Mperplexity: 16.73 -> 15.87Improved
replace GaLore-1/4 with FOAM-2GaLore · C4 LLaMA-60Mperplexity: 34.38 -> 28.53Improved
use FOAM-Mini instead of APOLLO-MiniAPOLLO-Mini · C4 LLaMA-60Mperplexity: 31.58 -> 29.71Improved
use FOAM-2 instead of APOLLO-1/4APOLLO · C4 LLaMA-7B at 150k stepsperplexity: 12.63 -> 11.13Improved
use FOAM instead of APOLLO for fine-tuningAPOLLO · MMLU Gemma3-1Baverage accuracy: 26.38 -> 25.99Hurt
use FOAM instead of Full-Adam for fine-tuningFull-Adam · GLUE RoBERTa-Largeaverage score: 87.66 -> 87.64No clear effect
lengthen sequence to 1024 with FOAM-2 vs APOLLO-1/4APOLLO · C4 LLaMA-350M seq len 1024perplexity: 18.77 -> 17.75Improved
use FOAM-2 with int8 states vs 8-bit Adam8-bit Adam · C4 LLaMA-60Mperplexity: 30.78 -> 28.86Improved

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