Hallucination Detection via Internal States and Structured Reasoning Consistency in Large Language Models
This work addresses a critical issue for AI safety and reliability by improving hallucination detection in LLMs, though it is incremental as it builds on and unifies existing methods.
The paper tackles the problem of detecting hallucinations in Large Language Models by addressing a 'Detection Dilemma' where existing methods fail on either factual or logical inconsistencies, and introduces a unified framework that resolves this gap, achieving significant performance improvements over baselines in experiments across three benchmarks and two LLMs.
The detection of sophisticated hallucinations in Large Language Models (LLMs) is hampered by a ``Detection Dilemma'': methods probing internal states (Internal State Probing) excel at identifying factual inconsistencies but fail on logical fallacies, while those verifying externalized reasoning (Chain-of-Thought Verification) show the opposite behavior. This schism creates a task-dependent blind spot: Chain-of-Thought Verification fails on fact-intensive tasks like open-domain QA where reasoning is ungrounded, while Internal State Probing is ineffective on logic-intensive tasks like mathematical reasoning where models are confidently wrong. We resolve this with a unified framework that bridges this critical gap. However, unification is hindered by two fundamental challenges: the Signal Scarcity Barrier, as coarse symbolic reasoning chains lack signals directly comparable to fine-grained internal states, and the Representational Alignment Barrier, a deep-seated mismatch between their underlying semantic spaces. To overcome these, we introduce a multi-path reasoning mechanism to obtain more comparable, fine-grained signals, and a segment-aware temporalized cross-attention module to adaptively fuse these now-aligned representations, pinpointing subtle dissonances. Extensive experiments on three diverse benchmarks and two leading LLMs demonstrate that our framework consistently and significantly outperforms strong baselines. Our code is available: https://github.com/peach918/HalluDet.