A comprehensive new survey from Stanford researchers examines reasoning failures in Large Language Models, identifying fundamental architectural limitations that persist despite impressive capabilities. The study, published in TMLR 2026, introduces a novel framework categorizing reasoning into embodied versus non-embodied types, with the latter split between intuitive and formal reasoning.

The researchers classify failures into three categories: fundamental architectural flaws affecting all tasks, application-specific limitations in particular domains, and robustness issues causing inconsistent performance across minor variations. For each failure type, they analyze root causes and present mitigation strategies, providing the first systematic examination of these shortcomings. The team also released a GitHub repository collecting research works on LLM reasoning failures.

By unifying fragmented research efforts, this survey offers valuable insights for building more reliable reasoning capabilities. The findings suggest that while LLMs excel at many tasks, significant gaps remain in their ability to reason consistently and accurately, particularly in formal logic and embodied scenarios. The comprehensive categorization framework provides researchers with a structured approach to addressing these persistent challenges.