ICLR 2026 Orals

Sequences of Logits Reveal the Low Rank Structure of Language Models

Noah Golowich, Allen Liu, Abhishek Shetty

LLMs & Reasoning Thu, Apr 23 · 3:39 PM–3:49 PM · 201 A/B Avg rating: 7.33 (6–8)
Author-provided TL;DR

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

Abstract

A major problem in the study of large language models is to understand their inherent low-dimensional structure. We introduce an approach to study the low-dimensional structure of language models at a model-agnostic level: as sequential probabilistic models. We first empirically demonstrate that a wide range of modern language models exhibit low-rank structure: in particular, matrices built from the model's logits for varying sets of prompts and responses have low approximate rank. We then show that this low-rank structure can be leveraged for generation --- in particular, we can generate a response to a target prompt using a linear combination of the model's outputs on unrelated, or even nonsensical prompts.

On the theoretical front, we observe that studying the approximate rank of language models in the sense discussed above yields a simple universal abstraction whose theoretical predictions parallel our experiments. We then analyze the representation power of the abstraction and give provable learning guarantees.

One-sentence summary·Auto-generated by claude-haiku-4-5-20251001(?)

Extended logit matrices reveal low-rank structure of language models enabling linear generation from unrelated prompts.

Contributions·Auto-generated by claude-haiku-4-5-20251001(?)
  • Demonstrates wide range of modern LLMs exhibit low-rank structure in logit matrices across prompts and responses
  • Shows low-rank structure can be leveraged for generation via linear combinations of model outputs
  • Develops simple universal abstraction with theoretical predictions paralleling experiments
  • Analyzes representation power and provides provable learning guarantees
Methods used·Auto-generated by claude-haiku-4-5-20251001(?)
  • Low-rank matrix analysis
  • Spectral methods
Limitations (author-stated)·Auto-generated by claude-haiku-4-5-20251001(?)

Authors did not state explicit limitations.

Future work (author-stated)·Auto-generated by claude-haiku-4-5-20251001(?)
  • Better understand how singular value decay evolves during training and use as diagnostic for training progress
    from the paper
  • Extract concepts and features from low-rank representation space for model-agnostic interpretability
    from the paper
  • Develop techniques to bypass safety guardrails using LINGEN and related frameworks
    from the paper
  • Explore safeguarding techniques against attacks suggested by the framework
    from the paper
  • Extend theoretical results to approximately low-rank models and investigate notions of approximation
    from the paper

Author keywords

  • Large language models
  • low-rank structure

Related orals

Something off? Let us know →