AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-22
Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory
The researchers used information theory to analyze when LLM-based monitors can detect attributes of outputs from reasoning traces and what limits their accuracy. They proposed a label-free training objective based on mutual information between outputs and chains of thought. Across multiple environments, it prevented chain-of-thought obfuscation even during training against a monitor and helped reduce reward hacking when task rewards were imperfectly specified.