Research map / Architectures and efficiency / Pruning and efficient tuning
Taming Momentum: Rethinking Optimizer States Through Low-Rank Approximation
Zhengbo Wang, Jian Liang, Ran He, Zilei Wang, Tieniu Tan
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 |
|---|---|---|---|
| use LoRA-Pre low-rank online momentum instead of Fira projection | Adam low-rank optimizer · C4 pre-training (Llama 350M) | perplexity: 17.22 -> 16.36 | Improved |
| use LoRA-Pre Adam instead of Fira at 60M | Adam low-rank optimizer · C4 pre-training (Llama 60M) | perplexity: 31.19 -> 32.57 | Hurt |
| use LoRA-Pre Adam instead of full-rank Adam | Llama 1B · C4 pre-training (Llama 1B) | perplexity: 15.56 -> 13.53 | Improved |
| use LoRA-Pre Adam instead of LoRA | Llama-2-7B · GSM8K (MetaMath100k fine-tune) | accuracy: 44.62 -> 57.35 | Improved |
| use LoRA-Pre Adam instead of LoRA | Llama-3.1-8B · MATH-500 (MetaMath100k fine-tune) | accuracy: 17.06 -> 17.66 | No clear effect |
| use LoRA-Pre Muon instead of LoRA Muon | Llama-2-7B · GSM8K (MetaMath100k fine-tune) | accuracy: 35.15 -> 47.20 | Improved |
| add momentum compression via LoRA-Pre to Muon instead of GaLore (ablation) | Muon low-rank · C4 pre-training (Llama 130M) | perplexity: 25.16 -> 23.05 | Improved |
| use projection-based low-rank Muon (GaLore) vs no momentum (ablation) | Muon · C4 pre-training (Llama 60M) | perplexity: 32.15 -> 34.39 | Hurt |
| increase rank (ablation) | LoRA-Pre and GaLore · C4 pre-training (Llama 60M/130M) | perplexity | Improved |
| use LoRA-Pre Adam at rank 16 instead of GaLore at rank 128 (ablation) | low-rank optimizer · C4 pre-training (Llama 60M) | perplexity | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- GaLore — GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (ICML 2024) · better on 6 of 6 shared results
- Fira — Fira: Can We Achieve Full-rank Training of LLMs Under Low-rank Constraint? (NeurIPS 2025) · better on 4 of 4 shared results
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
- Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning (NeurIPS 2026 · similarity 0.91)
- LoRA Training Provably Converges to a Low-Rank Global Minimum Or It Fails Loudly (But it Probably Won't Fail) (ICML 2025 · similarity 0.90)
- Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation (NeurIPS 2026 · similarity 0.90)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.90)
- LoRA-One: One-Step Full Gradient Could Suffice for Fine-Tuning Large Language Models, Provably and Efficiently (ICML 2025 · similarity 0.90)
- Breaking the Frozen Subspace: Importance Sampling for Low-Rank Optimization in LLM Pretraining (NeurIPS 2025 · similarity 0.90)
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.
