Research map / Large language models / Instruction and fine-tuning
WSM: Decay-Free Learning Rate Schedule via Checkpoint Merging for LLM Pre-training
Changxin Tian, jiapeng wang, Qian Zhao, Kunlong Chen, Jia Liu, Ziqi Liu, Jiaxin Mao, Xin Zhao, Zhiqiang Zhang, JUN ZHOU
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 |
|---|---|---|---|
| replace learning-rate decay with checkpoint merging (WSM mean merge) | WSD-annealed Ling-mini · Base model overall average | average score: 62.67 -> 63.95 | Improved |
| replace learning-rate decay with checkpoint merging, then SFT | WSD-annealed Ling-mini · Instruct overall average | average score: 62.90 -> 64.07 | Improved |
| replace learning-rate decay with checkpoint merging on code | SFT model · Instruct Code average | average score: 58.23 -> 57.95 | Hurt |
| replace learning-rate decay with checkpoint merging on professional knowledge | Base model · Professional Knowledge average | average score: 53.46 -> 56.04 | Improved |
| use mean merge instead of EMA (ablation) | WSM · Base overall average | average score: 63.01 -> 63.95 | Improved |
| use 1-sqrt merge weights instead of EMA (ablation) | WSM · Base overall average | average score: 63.01 -> 64.06 | Improved |
| use 1-sqrt merge weights instead of mean (ablation) | WSM · Base overall average | average score: 63.95 -> 64.06 | Improved |
| merge checkpoints instead of 1-sqrt LR decay (ablation) | WSD · Base overall average | average score: 62.67 -> 64.06 | Improved |
| finer granularity (5B,16) instead of single checkpoint (80B,1) (ablation) | WSM 80B merge · Base overall average | average score: 60.33 -> 63.63 | Improved |
| save every 10B and merge 8 instead of every 20B merge 4 (ablation) | WSM 80B merge · Base overall average | average score: 63.36 -> 63.78 | Improved |
| larger merge window (ablation) | WSM mean and 1-sqrt · Base benchmarks | average score: gains diminish as window grows | Mixed |
| combine decay and merging (Decay-then-Merge or Merge-then-Decay) (ablation) | WSM · Pre-training milestones | benchmark score | No clear effect |
Closest papers by meaning
- Maximizing Intermediate Checkpoint Value in LLM Pretraining with Bayesian Optimization (ICML 2025 · similarity 0.89)
- Straight to Zero: Why Linearly Decaying the Learning Rate to Zero Works Best for LLMs (ICLR 2025 · similarity 0.89)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.88)
- A Multi-Power Law for Loss Curve Prediction Across Learning Rate Schedules (ICLR 2025 · similarity 0.87)
- Power Lines: Scaling laws for weight decay and batch size in LLM pre-training (NeurIPS 2025 · similarity 0.87)
- Scaling Optimal LR Across Token Horizons (ICLR 2025 · similarity 0.87)
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.
