Two Pathways to Truthfulness: On the Intrinsic Encoding of LLM Hallucinations
This work addresses the issue of hallucinations in LLMs for users seeking more reliable AI systems, though it is incremental as it builds on prior findings about internal truthfulness signals.
The paper tackled the problem of understanding how large language models internally encode truthfulness to address hallucinations, identifying two distinct information pathways and proposing applications that enhance hallucination detection performance.
Despite their impressive capabilities, large language models (LLMs) frequently generate hallucinations. Previous work shows that their internal states encode rich signals of truthfulness, yet the origins and mechanisms of these signals remain unclear. In this paper, we demonstrate that truthfulness cues arise from two distinct information pathways: (1) a Question-Anchored pathway that depends on question-answer information flow, and (2) an Answer-Anchored pathway that derives self-contained evidence from the generated answer itself. First, we validate and disentangle these pathways through attention knockout and token patching. Afterwards, we uncover notable and intriguing properties of these two mechanisms. Further experiments reveal that (1) the two mechanisms are closely associated with LLM knowledge boundaries; and (2) internal representations are aware of their distinctions. Finally, building on these insightful findings, two applications are proposed to enhance hallucination detection performance. Overall, our work provides new insight into how LLMs internally encode truthfulness, offering directions for more reliable and self-aware generative systems.