Research map / Optimization / Optimizers
FOAM: Blocked State Folding for Memory-Efficient LLM Training
Ziqing Wen, Jiahuan Wang, Ping Luo, Dongsheng Li, Tao Sun
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add residual correction term R_t to blocked-average update (ablation) | FOAM · C4 pre-training LLaMA-60M/130M | validation perplexity | Improved |
| use FOAM-3 (level 3) instead of FOAM-2 (ablation) | FOAM · C4 LLaMA-1.3B | perplexity: 13.13 -> 13.19; memory 4.45G -> 3.97G | Mixed |
| vary level l (higher means less memory) (ablation) | FOAM · C4 pre-training | validation perplexity | No clear effect |
| replace Full-Adam with FOAM-2 | Full-Adam · C4 LLaMA-1.3B | perplexity: 14.51 -> 13.13 | Improved |
| replace APOLLO-1/4 with FOAM-2 | APOLLO · C4 LLaMA-350M | perplexity: 16.73 -> 15.87 | Improved |
| replace GaLore-1/4 with FOAM-2 | GaLore · C4 LLaMA-60M | perplexity: 34.38 -> 28.53 | Improved |
| use FOAM-Mini instead of APOLLO-Mini | APOLLO-Mini · C4 LLaMA-60M | perplexity: 31.58 -> 29.71 | Improved |
| use FOAM-2 instead of APOLLO-1/4 | APOLLO · C4 LLaMA-7B at 150k steps | perplexity: 12.63 -> 11.13 | Improved |
| use FOAM instead of APOLLO for fine-tuning | APOLLO · MMLU Gemma3-1B | average accuracy: 26.38 -> 25.99 | Hurt |
| use FOAM instead of Full-Adam for fine-tuning | Full-Adam · GLUE RoBERTa-Large | average score: 87.66 -> 87.64 | No clear effect |
| lengthen sequence to 1024 with FOAM-2 vs APOLLO-1/4 | APOLLO · C4 LLaMA-350M seq len 1024 | perplexity: 18.77 -> 17.75 | Improved |
| use FOAM-2 with int8 states vs 8-bit Adam | 8-bit Adam · C4 LLaMA-60M | perplexity: 30.78 -> 28.86 | Improved |
Closest papers by meaning
- A Memory Efficient Randomized Subspace Optimization Method for Training Large Language Models (ICML 2025 · similarity 0.89)
- Benchmarking Optimizers for Large Language Model Pretraining (NeurIPS 2026 · similarity 0.88)
- SPAM: Spike-Aware Adam with Momentum Reset for Stable LLM Training (ICLR 2025 · similarity 0.88)
- Mini-batch Coresets for Memory-efficient Language Model Training on Data Mixtures (ICLR 2025 · similarity 0.88)
- AdaRankGrad: Adaptive Gradient Rank and Moments for Memory-Efficient LLMs Training and Fine-Tuning (ICLR 2025 · similarity 0.88)
- Sophia: A Scalable Stochastic Second-order Optimizer for Language Model Pre-training (ICLR 2024 · similarity 0.88)
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.
