AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Pre-training under infinite compute

Konwoo Kim, Suhas Kotha, Percy Liang, Tatsunori Hashimoto

ICLR 2026 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws · 30 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
tune weight decay about 30x larger than default and jointly tune lr and epochsstandard recipe 200M tokens · DCLM validation lossloss asymptote: 3.75 -> 3.43Improved
increase epoch count too far300M model, 200M tokens · DCLM validation losslossHurt
increase parameter count to 1.4B without regularizationstandard recipe · DCLM validation losslossHurt
ensemble K models instead of scaling parametersregularized recipe · DCLM validation lossloss asymptote: 3.43 -> 3.34Improved
tune member hyperparameters for K to infinity (more epochs, less weight decay) (ablation)ensemble of 300M models · DCLM validation lossloss asymptote: 3.34 -> 3.27Improved
compose ensembling and parameter scaling (joint scaling)ensemble recipe · DCLM validation lossloss asymptote: 3.43 -> 3.17Improved
fix weight decay at 0.8 tuned only at 150M and transfer to larger models (ablation)regularized scaling · DCLM validation lossscaling behaviorHurt
fix epoch count 16 tuned only at 150M (ablation)regularized scaling · DCLM validation lossscaling behaviorHurt
distill 8-ensemble into 300M studentregularized 300M model · DCLM validation lossloss: 3.57 -> 3.36Improved
self-distill 300M teacher into same-size student300M model · DCLM validation losslossImproved
reduce batch size in continued pre-training (ablation)CPT Llama 3B · MegaMath CPT average accuracyaccuracy: 30.59 -> 34.48Improved
add epoching (K=1) to lower batch size (ablation)CPT Llama 3B · MegaMath CPT average accuracyaccuracy: 34.48 -> 35.82Improved

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