Correct Chains, Wrong Answers: Dissociating Reasoning from Output in LLM Logic

arXiv:2604.13065v1 Announce Type: cross Abstract: LLMs can execute every step of chain-of-thought reasoning correctly and still produce wrong final answers. We introduce the Novel Operator Test, a benchmark that separates operator logic from operator name, enabling rigorous distinction between...