Dissecting Failure Dynamics in Large Language Model Reasoning
arXiv:2604.14528v1 Announce Type: cross
Abstract: Large Language Models (LLMs) achieve strong performance through extended inference-time deliberation, yet how their reasoning failures arise remains poorly understood. By analyzing model-generated reas…