Research map / Optimization / Optimizers
Memory-Efficient LLM Pretraining via Minimalist Optimizer Design
Athanasios Glentis, Jiaxiang Li, Andi Han, Mingyi Hong
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 last-layer momentum to column-wise SGD (ablation) | SGD column-wise · C4 LLaMA-130M | perplexity: 28.85 -> 22.57 | Improved |
| add last-layer momentum to column-wise SGD (ablation) | SGD column-wise · C4 LLaMA-60M | perplexity: 39.89 -> 30.81 | Improved |
| add last-layer momentum to column-wise SGD (ablation) | SGD column-wise · C4 LLaMA-350M | perplexity: 20.38 -> 16.32 | Improved |
| add first-layer momentum on top of last-layer momentum (ablation) | SCALE · C4 LLaMA-130M | perplexity: 22.57 -> 22.26 | No clear effect |
| column-wise instead of row-wise normalization (ablation) | SGD · C4 LLaMA-130M | perplexity: 37.67 -> 28.85 | Improved |
| column-wise instead of sign normalization (ablation) | SGD · C4 LLaMA-130M | perplexity: 40.42 -> 28.85 | Improved |
| singular-value (NS) instead of column-wise normalization, no momentum (ablation) | SGD · C4 LLaMA-130M | perplexity: 25.25 vs 28.85 | Hurt |
| column-wise instead of row-wise normalization on last layer (ablation) | SCALE · C4 LLaMA-130M | perplexity: 28.83 -> 22.57 | Improved |
| column-wise instead of row-wise normalization on non-last layers (ablation) | SCALE · C4 LLaMA-130M | perplexity: 23.27 -> 22.57 | Improved |
| column-wise instead of row-wise normalization on first layer (ablation) | SCALE · C4 LLaMA-130M | perplexity: 22.94 -> 22.57 | Improved |
| column-wise instead of normalizing along larger dimension (ablation) | SCALE · C4 LLaMA-130M | perplexity: 23.52 -> 22.57 | Improved |
| singular-value (NS) instead of column-wise with last-layer momentum (ablation) | SCALE · C4 LLaMA-350M | perplexity: 16.67 vs 16.32 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SWAN — SWAN: SGD with Normalization and Whitening Enables Stateless LLM Training (ICML 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Benchmarking Optimizers for Large Language Model Pretraining (NeurIPS 2026 · similarity 0.92)
- SWAN: SGD with Normalization and Whitening Enables Stateless LLM Training (ICML 2025 · similarity 0.92)
- Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark (ICML 2024 · similarity 0.91)
- Fira: Can We Achieve Full-rank Training of LLMs Under Low-rank Constraint? (NeurIPS 2025 · similarity 0.91)
- A Memory Efficient Randomized Subspace Optimization Method for Training Large Language Models (ICML 2025 · similarity 0.91)
- GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (ICML 2024 · similarity 0.91)
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.
