Research map / Architectures and efficiency / Pruning and efficient tuning
One LR Doesn’t Fit All: Heavy-Tail Guided Layerwise Learning Rates for LLMs
Di He, Songjun Tu, Keyu Wang, Lu Yin, Shiwei Liu
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 Hill-alpha guided layerwise LR (LLR) | AdamW LLaMa pretraining · FineWeb 60M | validation perplexity: 21.94 -> 20.30 | Improved |
| add LLR | AdamW LLaMa-1B · FineWeb 1B | validation perplexity: 9.77 -> 9.59 | Improved |
| add LLR | LLaMa-1B · 7 zero-shot commonsense tasks | average accuracy (%): 47.09 -> 49.02 | Improved |
| add LLR at larger token budget (3B model, 30B tokens) | LLaMa-3B · zero-shot commonsense | average accuracy (%): 48.58 -> 50.61 | Improved |
| add LLR at 5x chinchilla tokens (100B) | LLaMa-1B · zero-shot commonsense | average accuracy (%): 51.16 -> 51.86 | Improved |
| add LLR | Llama-3.2-1B finetuning · Commonsense reasoning finetuning | average accuracy (%): 52.94 -> 53.58 | Improved |
| add LLR to ViT-Tiny | ViT-Tiny · ImageNet-1K | top-1 accuracy (%): 67.42 -> 68.51 | Improved |
| add LLR with Muon optimizer | LLaMa-350M 20B tokens · zero-shot average | average accuracy (%): 45.53 -> 47.08 | Improved |
| add LLR with AdamW at 3x chinchilla | LLaMa-350M 20B tokens · zero-shot average | average accuracy (%): 44.50 -> 46.03 | Improved |
| invert the Hill-alpha to LR mapping (LLR-inv) instead of the proposed one (ablation) | LLR · FineWeb 135M | validation perplexity: 17.03 -> 27.18 | Hurt |
| use linear instead of Log2 or Sqrt LR assignment (ablation) | LLR · LLaMa-60M, LR 0.0005 | validation perplexity: 17.86 (Log2) -> 17.55 (Linear) | Improved |
| fix embedding LR at upper bound instead of LLR-unembed (ablation) | LLR · LLaMa pretraining | validation perplexity | Improved |
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.
- Adammini — Adam-mini: Use Fewer Learning Rates To Gain More (ICLR 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
- Building on Efficient Foundations: Effective Training of LLMs with Structured Feedforward Layers (NeurIPS 2024 · similarity 0.90)
- MISA: Memory-Efficient LLMs Optimization with Module-wise Importance Sampling (NeurIPS 2025 · similarity 0.90)
- The Curse of Depth in Large Language Models (NeurIPS 2025 · similarity 0.89)
- Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment (NeurIPS 2024 · similarity 0.89)
- LISA: Layerwise Importance Sampling for Memory-Efficient Large Language Model Fine-Tuning (NeurIPS 2024 · similarity 0.89)
- Relaxed Recursive Transformers: Effective Parameter Sharing with Layer-wise LoRA (ICLR 2025 · similarity 0.89)
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.
