Research map / Architectures and efficiency / RNNs and normalization
Understanding the Mechanisms of Fast Hyperparameter Transfer
Nikhil Ghosh, Denny Wu, Alberto Bietti
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 |
|---|---|---|---|
| transfer the width-128 optimal learning rate to larger widths with Adam | Llama with muP on WikiText-103 · WikiText-103 | EMA loss | No clear effect |
| switch the optimizer from Adam to Muon | Llama with muP on WikiText-103 · WikiText-103 | LR transfer quality: top-k invariance up to k ≈ 60 for Adam vs k_0 ≈ 10 for Muon | Hurt |
| bound Dion rank at min(n/2, 128) instead of n/2 (ablation) | Dion optimizer · WikiText-103 Llama | LR transfer stability: rank min(n/2, 128) vs n/2 | Improved |
| evaluate LR sweeps on easy subset (lowest 25% MCI) instead of full validation set (ablation) | two-layer MLP on CIFAR-10 · CIFAR-10 validation | LR transfer: lowest 25% MCI | Improved |
| evaluate LR sweeps on hard subset (highest 5% MCI) (ablation) | two-layer MLP on CIFAR-10 · CIFAR-10 validation | LR transfer: highest 5% MCI | Hurt |
| use EMA-smoothed trajectory instead of the base trajectory (ablation) | trajectory linearization · WikiText-103 | final loss | No clear effect |
| scale width of a two-layer ReLU network under muP to learn a norm indicator | two-layer ReLU network · norm indicator, d=2^6, n=2^14 | fast HP transfer: b_n ~ sqrt(a_n) | No clear effect |
Closest papers by meaning
- On the Provable Separation of Scales in Maximal Update Parameterization (ICML 2025 · similarity 0.91)
- Power Lines: Scaling laws for weight decay and batch size in LLM pre-training (NeurIPS 2025 · similarity 0.89)
- Scaling Optimal LR Across Token Horizons (ICLR 2025 · similarity 0.88)
- Depthwise Hyperparameter Transfer in Residual Networks: Dynamics and Scaling Limit (ICLR 2024 · similarity 0.88)
- Scaling Exponents Across Parameterizations and Optimizers (ICML 2024 · similarity 0.87)
- Don't be lazy: CompleteP enables compute-efficient deep transformers (NeurIPS 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.
