![[2602.23163] A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.23163] A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring
Summary
This paper presents a decision-theoretic framework for understanding steganography in large language models (LLMs), addressing the challenges of detecting and quantifying hidden information in these systems.
Why It Matters
As LLMs become more prevalent, their potential for misalignment and hidden communication poses significant risks. This research offers a novel approach to detect and mitigate such risks, enhancing oversight mechanisms in AI systems.
Key Takeaways
- Introduces a decision-theoretic perspective on steganography in LLMs.
- Proposes a new measure, the steganographic gap, to quantify hidden information.
- Validates the framework empirically, demonstrating its utility in monitoring LLM behavior.
Computer Science > Artificial Intelligence arXiv:2602.23163 (cs) [Submitted on 26 Feb 2026] Title:A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring Authors:Usman Anwar, Julianna Piskorz, David D. Baek, David Africa, Jim Weatherall, Max Tegmark, Christian Schroeder de Witt, Mihaela van der Schaar, David Krueger View a PDF of the paper titled A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring, by Usman Anwar and 8 other authors View PDF HTML (experimental) Abstract:Large language models are beginning to show steganographic capabilities. Such capabilities could allow misaligned models to evade oversight mechanisms. Yet principled methods to detect and quantify such behaviours are lacking. Classical definitions of steganography, and detection methods based on them, require a known reference distribution of non-steganographic signals. For the case of steganographic reasoning in LLMs, knowing such a reference distribution is not feasible; this renders these approaches inapplicable. We propose an alternative, \textbf{decision-theoretic view of steganography}. Our central insight is that steganography creates an asymmetry in usable information between agents who can and cannot decode the hidden content (present within a steganographic signal), and this otherwise latent asymmetry can be inferred from the agents' observable actions. To formalise this perspective, we introduce generalised $\mathcal{V}$-informa...