Reasoning Trace Inversion: Detecting When LLMs Answer the Wrong Question Instead of Abstaining

arXiv·medium signal

Researchers from IBM propose a Query Misalignment Framework showing that reasoning model hallucinations can be reinterpreted as the LLM answering a different question than the one asked. By inverting the reasoning trace to reconstruct what question the model actually answered, the framework detects misalignment between query and response — a practical signal for abstention in deployed reasoning systems where current models have notably worse abstention abilities.