Sequences of Logits Reveal the Low Rank Structure of Language Models

ICLROral2026

Authors
Noah Golowich, Allen Liu, Abhishek Shetty
Affiliation
Microsoft
Venue
ICLR 2026
Track
Oral

TL;DR

We exploit the low-rank structure of the logit matrices of LLMs to draw new empirical and theoretical conclusions.

Opening excerpt from the authors’ abstract. source

Read the paper

Topics

language model llm

← All ICLR 2026 Oral papers · Browse the whole archive