What are failure modes of large language models in reasoning tasks?

LLMs fail due to hallucination, weak compositionality, and brittle multi-step reasoning.

LLMs often fail when reasoning requires strict logical consistency or long dependency chains. Errors accumulate across steps, and probabilistic token generation may diverge from valid logical paths. These failures are structural, not just data-related.