Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory
The work establishes a formal information-theoretic framework to measure and improve how reliably Chain-of-Thought (CoT) traces can detect undesirable behaviors like test-hacking in LLMs. It proves that mere correlation is insufficient, introducing concepts like Information Gap and Elicitation Error to quantify monitoring failures. The paper proposes training objectives to systematically enhance monitor performance.