Decision-Theoretic Steganography for LLM Monitoring

A recent paper on arXiv presents a decision-theoretic model aimed at identifying steganography within large language models (LLMs). Traditional definitions of steganography necessitate a known reference distribution of signals that do not contain hidden information, which poses challenges for LLM reasoning. The authors propose a generalized V-information method that utilizes the disparity in accessible information between agents capable of decoding hidden messages and those that are not. This disparity can be deduced from observable behaviors, allowing for detection without needing a reference distribution. The research tackles the potential risk of misaligned models employing steganography to circumvent monitoring, providing a systematic approach to detect and measure such actions.