Compressed Sensing for Capability Localization in Large Language Models

Anna Bair, Yixuan Even Xu, Mingjie Sun, J. Zico Kolter · 2026

Read Compressed Sensing for Capability Localization in Large Language Models by Anna Bair, Yixuan Even Xu, Mingjie Sun, J. Zico Kolter (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-03335.txt.

Description

Anna Bair and three colleagues show that single skills inside a language model live in surprisingly few parts. Models reason about mathematics, write code, and handle language, and one might expect each skill to be spread everywhere. Their 2026 study finds the opposite: many capabilities sit in small subsets of attention heads, the components that route information between tokens.

Turning off as few as five skill-specific heads can cut performance on that skill's tests by up to about two thirds, while unrelated tasks stay mostly intact. Finding the right heads is the hard part, so the authors borrow compressed sensing, using strategic knockouts and few evaluations to locate them. Tests across two model families at several sizes confirm a modular pattern, with clear stakes for interpretation, editing, and safety work.

Catalog record

CREATORAnna Bair, Yixuan Even Xu, Mingjie Sun, J. Zico Kolter
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44460
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.03335

Browse the full library · About the Tower of Babel