Research map / Architectures and efficiency / RNNs and normalization
$\mu$pscaling small models: Principled warm starts and hyperparameter transfer
Yuxin Ma, Nan Chen, Mateo Diaz Diaz, Soufiane Hayou, Dmitriy Kunisky, Soledad Villar
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 |
|---|---|---|---|
| upscale a trained narrow model with noise injection and muP hyperparameters instead of training the wide model from scratch | MLP, ResNet-18, GPT-2 · Forest Cover Type, CIFAR-100, FineWeb (training loss) | training loss | Improved |
| upscale instead of training from scratch, measured on validation | ResNet-18 · CIFAR-100 validation | validation performance | Hurt |
| upscale instead of training from scratch, measured on validation | MLP and GPT-2 · Forest Cover Type, FineWeb validation | validation performance | Improved |
| tune noise and learning rate at small width and transfer, instead of tuning at the target width as in Net2Net | upscaling procedure · GPT-2 FineWeb, n/n0=10 | per-run FLOP cost: 48.2x speedup | Improved |
| inject nonzero symmetry-breaking noise into the widened model versus zero noise (ablation) | upscaling procedure · all three settings | initialization and terminal loss | Improved |
| transfer optimal learning rate and noise level across target widths 128 to 2048 (ablation) | upscaled GPT-2 and MLP · GPT-2 AdamW, MLP SGD, MLP AdamW | optimal learning rate and noise level | No clear effect |
| use zero noise with muP scaling (ablation) | widened model · architectures and optimizers in appendix | training trajectory | No clear effect |
Closest papers by meaning
- On the Provable Separation of Scales in Maximal Update Parameterization (ICML 2025 · similarity 0.90)
- Depthwise Hyperparameter Transfer in Residual Networks: Dynamics and Scaling Limit (ICLR 2024 · similarity 0.88)
- Scaling Exponents Across Parameterizations and Optimizers (ICML 2024 · similarity 0.88)
- u-$\mu$P: The Unit-Scaled Maximal Update Parametrization (ICLR 2025 · similarity 0.88)
- $\boldsymbol{\mu}\mathbf{P^2}$: Effective Sharpness Aware Minimization Requires Layerwise Perturbation Scaling (NeurIPS 2024 · similarity 0.87)
- µnit Scaling: Simple and Scalable FP8 LLM Training (ICML 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.
