AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Taming Curvature: Architecture Warm-up for Stable Transformer Training

Sameera Ramasinghe, Thalaiyasingam Ajanthan, Hadi Mohaghegh Dolatabadi, Chamin P Hewa Koneputugodage, Gil Avraham, Violetta Shevchenko, Yan Zuo, Karol Pajak, Alexander Long

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Transformer block and depth design · 0 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
add architecture warm-up (start at half depth, unlock layers gradually)Llama-style 1B Transformer · FineWeb perplexity (peak LR 8e-3)validation perplexity: QK-Norm 49.88 -> 25.02; baseline divergedImproved
add architecture warm-upLlama-style 1B Transformer · OLMo-Mix perplexityvalidation perplexity: all baselines diverged -> 18.54Improved
add architecture warm-up versus SoftcapLlama-style 1B Transformer · DCLM perplexityvalidation perplexity: 43.38 -> 22.64Improved
use QK-Clip stabilizationLlama-style 1B Transformer · FineWeb/DCLM/OLMo-MixdivergenceHurt
architecture warm-up vs QK-Norm at Chinchilla-optimal compute1B, 25B tokens · FineWebvalidation perplexity: 20.28 -> 18.35Improved
increase depth from 8 to 16/32 layers at LR 8e-3 without warm-up (ablation)Transformer · FineWeb curvature and stabilitylambda_max and loss spikesHurt
remove LR warmup and use architecture warm-up only (ablation)1B Transformer · FineWebconvergence and stabilityNo clear effect
change unlock spacing of layers to zero (ablation)Architecture warm-up · FineWebconvergenceNo clear effect
warm-start the power iteration with previous top eigenvector (ablation)online Hessian eigenvalue estimation · 4-layer Transformerrelative error of curvatureImproved
sweep peak LR 3e-3 to 8e-3 with architecture warm-up (ablation)Transformer · FineWebstabilityImproved

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