AI Paper Atlas

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

Sequences of Logits Reveal the Low Rank Structure of Language Models

Noah Golowich, Allen Liu, Abhishek Shetty

ICLR 2026 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data

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
train the model (vs. untrained step-0 checkpoint) (ablation)OLMo-1b · wiki logit matrixsingular value power-law exponent: 0.374 -> 0.561Improved
use extended logit matrix (multi-token futures) instead of single-token logit matrix for generation (ablation)LINGEN · OLMo-1b generationKL divergence to true modelImproved
generate using linear combination of model outputs on in-distribution historiesLINGEN · OLMo-1b, wikiKL divergence per tokenImproved
generate using nonsense histories and futuresLINGEN · OLMo-1b, out-of-distributionKL divergence per tokenMixed
compare against earlier training checkpoint baselineLINGEN · OLMo-1bKL divergenceImproved
increase k (columns) from 50 to 200 or choose random tokens (ablation)logit matrix · OLMo-1bpower-law exponent: 0.561 -> 0.543 / 0.527No clear effect
scale up size of histories and futures (ablation)logit matrix · OLMo-1bpower-law exponent alphaNo clear effect
project futures to random-token (nonsense) futurescolumn space of low-rank approximation · OLMo-1bprincipal angle cosinesNo 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