AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Understanding the Mechanisms of Fast Hyperparameter Transfer

Nikhil Ghosh, Denny Wu, Alberto Bietti

ICLR 2026 · Architectures and efficiency › RNNs and normalization · approach: Dataset distillation and data-efficient training · 6 citations (OpenAlex)

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 changedWhereMeasuredEffect
transfer the width-128 optimal learning rate to larger widths with AdamLlama with muP on WikiText-103 · WikiText-103EMA lossNo clear effect
switch the optimizer from Adam to MuonLlama with muP on WikiText-103 · WikiText-103LR transfer quality: top-k invariance up to k ≈ 60 for Adam vs k_0 ≈ 10 for MuonHurt
bound Dion rank at min(n/2, 128) instead of n/2 (ablation)Dion optimizer · WikiText-103 LlamaLR transfer stability: rank min(n/2, 128) vs n/2Improved
evaluate LR sweeps on easy subset (lowest 25% MCI) instead of full validation set (ablation)two-layer MLP on CIFAR-10 · CIFAR-10 validationLR transfer: lowest 25% MCIImproved
evaluate LR sweeps on hard subset (highest 5% MCI) (ablation)two-layer MLP on CIFAR-10 · CIFAR-10 validationLR transfer: highest 5% MCIHurt
use EMA-smoothed trajectory instead of the base trajectory (ablation)trajectory linearization · WikiText-103final lossNo clear effect
scale width of a two-layer ReLU network under muP to learn a norm indicatortwo-layer ReLU network · norm indicator, d=2^6, n=2^14fast HP transfer: b_n ~ sqrt(a_n)No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy