Robots Atlas>ROBOTS ATLAS

AI Agent Architecture — ReAct, Memory, Planning and Multi-Agent Systems · Reasoning — How an LLM Thinks Before Acting

When Reasoning Fails — CoT Pitfalls and Limits of Self-Correction

Reasoning — How an LLM Thinks Before Acting

Introduction

Chain-of-Thought and Tree of Thoughts are not infallible — models make characteristic errors: they fabricate convincing rationale for wrong conclusions, fall victim to composition errors, struggle with negation, and cannot effectively correct their own mistakes without an external signal. This lesson catalogs the most important CoT pitfalls, the phenomenon of "unfaithful reasoning", and the limits of self-correction — so developers know when reasoning is a risky tool.