Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations

2025-04-17

AI Summary

Large language models (LLMs) frequently generate hallucinated content inconsistent with facts or context, a phenomenon whose complex, multifaceted origins remain difficult to diagnose. This paper introduces the *subsequence association causal explanation paradigm*, the first framework to attribute hallucinations to dominant erroneous subsequence associations in the decoder that override faithful ones, and formally establishes that Transformer decoders intrinsically function as subsequence embedding models. Methodologically, we propose a randomized-context probabilistic tracing algorithm, integrated with linear-layer association analysis and empirical validation via alignment with training corpus evidence—yielding interpretable, reproducible hallucination溯源. Experiments across multiple tasks demonstrate that our approach significantly outperforms standard attribution baselines: it precisely localizes causal hallucinatory subsequences, and the identified associations exhibit strong statistical alignment with their actual distributions in the training corpus.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

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.

Problem

• Understanding why LLMs generate hallucinatory content
• Tracing hallucinations via subsequence association framework
• Identifying causal subsequences causing model hallucinations

Innovation

• Subsequence association framework traces hallucinations systematically
• Decoder-only transformers act as subsequence embedding models
• Tracing algorithm identifies causal subsequences via hallucination probabilities