AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Hidden Breakthroughs in Language Model Training

Sara Kangaslahti, Elan Rosenfeld, Naomi Saphra

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes · 12 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
Cluster on POLCA decomposed loss instead of exact lossexact-loss clustering · 3-digit additionmax carry homogeneity: 0.514 -> 0.973Improved
Cluster on POLCA instead of parameter-aligned LCALCA · 3-digit additionmax carry homogeneity: 0.792 -> 0.973Improved
Use POLCA instead of exact loss for hidden breakthroughsexact-loss clustering · 3-digit additionfraction of clusters with hidden breakthroughs: 0.0 -> 0.355Improved
Add second-order term to POLCA (ablation)first-order POLCA · 3-digit additionmax carry homogeneity: 0.948 -> 0.973Improved
Add second-order term to POLCA (ablation)first-order POLCA · 3-digit additionF1 of carry cluster: 0.848 -> 0.850No clear effect
Use Hessian eigenvector basis instead of random orthonormal basis (ablation)POLCA · 3-digit additionfraction of clusters with hidden breakthroughs: 0.031 -> 0.355Improved
Use Hessian eigenvector basis instead of random orthonormal basis (ablation)POLCA · 3-digit additionmax carry homogeneity: 0.902 -> 0.973Improved
Use correctly ordered Hessian basis instead of shuffled checkpoints (ablation)POLCA · 3-digit additionmax carry homogeneity: 0.856 -> 0.973Improved
Cluster on projected Fisher information instead of POLCA (ablation)POLCA · 3-digit additionmax carry homogeneity: 0.973 -> 0.664Hurt
Add second-order approximation at datapoint level (ablation)POLCA · 3-digit additionRMSE of approximation: RMSE 0.145No 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