H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons

2025-12-01

AI Summary

Hallucination in large language models (LLMs) severely undermines their reliability, yet its underlying neural mechanisms remain poorly understood. This work systematically uncovers the micro-scale neural basis of hallucination generation at the neuron level. Through fine-grained neuron-level activation analysis, causal intervention experiments, and cross-task generalization tests, we identify a sparse set of neurons—constituting less than 0.1% of total neurons—that stably predict hallucinatory outputs across diverse tasks and model architectures. Crucially, these neurons are shown to emerge during pretraining and exhibit a causal relationship with model overcompliance behavior. Our study establishes, for the first time, an interpretable link between macroscopic hallucination phenomena and microscopic neural activity. This yields a novel, mechanistic paradigm for targeted intervention and provides actionable pathways toward enhancing LLM reliability through neuron-level steering.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Large language models (LLMs) frequently generate hallucinations -- plausible but factually incorrect outputs -- undermining their reliability. While prior work has examined hallucinations from macroscopic perspectives such as training data and objectives, the underlying neuron-level mechanisms remain largely unexplored. In this paper, we conduct a systematic investigation into hallucination-associated neurons (H-Neurons) in LLMs from three perspectives: identification, behavioral impact, and origins. Regarding their identification, we demonstrate that a remarkably sparse subset of neurons (less than $0.1%$ of total neurons) can reliably predict hallucination occurrences, with strong generalization across diverse scenarios. In terms of behavioral impact, controlled interventions reveal that these neurons are causally linked to over-compliance behaviors. Concerning their origins, we trace these neurons back to the pre-trained base models and find that these neurons remain predictive for hallucination detection, indicating they emerge during pre-training. Our findings bridge macroscopic behavioral patterns with microscopic neural mechanisms, offering insights for developing more reliable LLMs.

Problem

• Identifies sparse neurons predicting hallucinations in LLMs
• Shows these neurons cause over-compliance behaviors in models
• Traces hallucination neurons to origins in pre-training phase

Innovation

• Sparse neurons predict hallucinations accurately
• Neurons cause over-compliance behaviors directly
• Hallucination neurons originate in pre-training phase