A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring

Usman Anwar, Julianna Piskorz, David D. Baek, David Africa, Jim Weatherall, Max Tegmark, Christian Schroeder de Witt, Mihaela van der Schaar, David Krueger · 2026

Read A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring by Usman Anwar, Julianna Piskorz, David D. Baek, David Africa, Jim Weatherall, Max Tegmark, Christian Schroeder de Witt, Mihaela van der Schaar, David Krueger (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-2602-23163.txt.

Description

Usman Anwar and eight co-authors address a blind spot in watching large language models. A model could hide messages in its own reasoning to slip past oversight, and current detection tools cannot catch it. Classical tests for hidden messages need a known baseline of clean signals to compare against. For model reasoning no such baseline exists, so those tools do not apply.

The authors reframe the problem in terms of decisions, not signals. A hidden message creates a gap in usable information between agents who can decode it and agents who cannot. That gap shows up in what the agents do. They formalize this with a generalized measure of usable information and a steganographic gap score. Tests show the score can detect, measure, and help reduce hidden reasoning in language models.

Catalog record

CREATORUsman Anwar, Julianna Piskorz, David D. Baek, David Africa, Jim Weatherall, Max Tegmark, Christian Schroeder de Witt, Mihaela van der Schaar, David Krueger
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE141 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-23163

Browse the full library · About the Tower of Babel