Sanity Checks for Long-Form Hallucination Detection in LLMs

A new research paper titled 'Sanity Checks for Long-Form Hallucination Detection' addresses a critical gap in evaluating large language models (LLMs). While current hallucination detection methods increasingly rely on chain-of-thought reasoning traces, it remains unclear whether these methods assess the reasoning process itself or merely exploit surface-level correlates of the final answer. The authors introduce a controlled-invariance methodology featuring two oracle tests: 'Force,' which replaces the final answer with ground truth while keeping the reasoning trace, and 'Remove,' which strips answer-announcement steps. These tests reveal that many existing detectors rely on answer-level artifacts rather than valid intermediate reasoning. The study further presents TRACT, a lightweight scorer using lexical trajectory features like hedging trends and step-length dynamics. TRACT demonstrates strong robustness and competes with or outperforms complex baselines once artifacts are controlled. The findings suggest that the primary challenge in reasoning-aware hallucination detection is isolating signal from endpoint cues, not a lack of signal within the reasoning trace itself.