Two Pathways to Truthfulness: On the Intrinsic Encoding of LLM Hallucinations

TL;DR

This paper uncovers two distinct internal pathways in large language models that encode signals of truthfulness, providing insights for improving hallucination detection and model reliability.

Contribution

It identifies and disentangles two mechanisms of truthfulness encoding in LLMs, revealing their properties and potential for enhancing hallucination detection.

Findings

Two pathways are identified: Question-Anchored and Answer-Anchored.

Disentanglement of pathways through attention knockout and token patching.

Pathways are linked to LLM knowledge boundaries and are internally recognized.

Abstract

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.