Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Sun, Yiyou, Gai, Yu, Chen, Lijie, Ravichander, Abhilasha, Choi, Yejin, Song, Dawn
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2504.12691
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866915247762178048
author Sun, Yiyou
Gai, Yu
Chen, Lijie
Ravichander, Abhilasha
Choi, Yejin
Song, Dawn
author_facet Sun, Yiyou
Gai, Yu
Chen, Lijie
Ravichander, Abhilasha
Choi, Yejin
Song, Dawn
contents Large language models (LLMs) frequently generate hallucinations-content that deviates from factual accuracy or provided context-posing challenges for diagnosis due to the complex interplay of underlying causes. This paper introduces a subsequence association framework to systematically trace and understand hallucinations. Our key insight is that hallucinations arise when dominant hallucinatory associations outweigh faithful ones. Through theoretical and empirical analyses, we demonstrate that decoder-only transformers effectively function as subsequence embedding models, with linear layers encoding input-output associations. We propose a tracing algorithm that identifies causal subsequences by analyzing hallucination probabilities across randomized input contexts. Experiments show our method outperforms standard attribution techniques in identifying hallucination causes and aligns with evidence from the model's training corpus. This work provides a unified perspective on hallucinations and a robust framework for their tracing and analysis.
format Preprint
id arxiv_https___arxiv_org_abs_2504_12691
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations
Sun, Yiyou
Gai, Yu
Chen, Lijie
Ravichander, Abhilasha
Choi, Yejin
Song, Dawn
Computation and Language
Large language models (LLMs) frequently generate hallucinations-content that deviates from factual accuracy or provided context-posing challenges for diagnosis due to the complex interplay of underlying causes. This paper introduces a subsequence association framework to systematically trace and understand hallucinations. Our key insight is that hallucinations arise when dominant hallucinatory associations outweigh faithful ones. Through theoretical and empirical analyses, we demonstrate that decoder-only transformers effectively function as subsequence embedding models, with linear layers encoding input-output associations. We propose a tracing algorithm that identifies causal subsequences by analyzing hallucination probabilities across randomized input contexts. Experiments show our method outperforms standard attribution techniques in identifying hallucination causes and aligns with evidence from the model's training corpus. This work provides a unified perspective on hallucinations and a robust framework for their tracing and analysis.
title Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations
topic Computation and Language
url https://arxiv.org/abs/2504.12691