Information-Theoretic Analysis of Chain-of-Thought Monitorability
October 22, 2026
An information-theoretic framework identifies information gap and elicitation error as the primary drivers of failure in LLM-based Chain-of-Thought (CoT) monitors. The study shows that non-zero mutual information between CoT and output is necessary but insufficient for effective detection of reasoning-based issues like test-hacking.
HOW THIS AFFECTS YOU
●
researcherYou can use these metrics to quantify why your LLM monitoring systems are failing to detect reasoning errors.
●
policyThis highlights why relying on reasoning traces for safety audits may be technically insufficient.