AI & ML Nature Is Weird

LLMs can perform every single logical step in a reasoning chain perfectly and still confidently hallucinate the wrong final answer.

April 16, 2026

Original Paper

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

Abinav Rao, Sujan Rachuri, Nikhil Vemuri

arXiv · 2604.13065

The Takeaway

We've long relied on 'Chain of Thought' (CoT) as a proxy for model reliability, but this paper proves that 'correct reasoning' and 'correct output' are actually dissociated. The model can execute a flawless internal logic trace only to fail at the final integration step. This breaks the assumption that if the logic is right, the result must be right. It reveals a deep architectural flaw where the 'reasoning engine' and the 'answer generator' are disconnected. For developers, this means CoT is not enough for verification; you need to build secondary checkers that specifically audit the 'landing' of a logic chain. This is a wake-up call for anyone building high-stakes logic-dependent AI systems.

From the 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 genuine reasoning and pattern retrieval. By evaluating Boolean operators under unfamiliar names across depths 1-10 on five models (up to 8,100 problems each), we demonstrate a reasoning-output dissociation that existing benchmarks cannot detect. At Claude