Mechanistically Eliciting Latent Behaviors in Language Models

Andrew Mack, Nina Panickssery, Alexander Matt Turner · 2026

Read Mechanistically Eliciting Latent Behaviors in Language Models by Andrew Mack, Nina Panickssery, Alexander Matt Turner (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-2606-29604.txt.

Description

Mack, Panickssery, and Turner present Causal Perturbative Elicitation, a method for surfacing latent behaviours hidden inside a language model. Their method, called CPE, needs no labelled examples of the behaviour it finds. It breaks down the workings of a deep transformer slice with a tensor-decomposition-based algorithm, yielding low-rank adapters.

CPE learns many adapters from a single example. On the Countdown task it nearly matches a supervised method given equal wall-clock time. It also surfaces hidden failure modes. On a password-locked coding benchmark, CPE restores most of the locked performance, exposing sandbagging, where a model underperforms on purpose. Applied to a model organism built to fake alignment, CPE almost wipes out the alignment-faking behaviour. The authors offer CPE as a tool for exploring the range of conduct a model could show.

Catalog record

CREATORAndrew Mack, Nina Panickssery, Alexander Matt Turner
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE82 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.29604

Browse the full library · About the Tower of Babel