Detecting Reasoning Failures in LLMs via Chain-of-Thought Dynamics

A new research paper on arXiv (2608.03291) proposes a method to detect reasoning failures in large language models (LLMs) by analyzing the dynamics of chain-of-thought (CoT) traces. The authors argue that existing approaches focus on semantic correctness of individual steps, missing failures distributed across the reasoning trajectory. They also note that verbalized CoTs may not faithfully reflect internal reasoning, so their method does not assume semantic faithfulness. Instead, they ask whether the dynamics of visible CoT can systematically distinguish successful from failed reasoning. The paper is categorized as a cross announcement and is available at the provided URL.