Do Hallucination Neurons Generalize? Evidence from Cross-Domain Transfer in LLMs

TL;DR

This study investigates whether hallucination neurons in large language models generalize across different knowledge domains, finding they do not, which impacts how hallucination detection should be implemented.

Contribution

It provides evidence that hallucination neurons are domain-specific, challenging the idea of a universal neural signature for hallucinations in LLMs.

Findings

Hallucination neurons do not generalize well across domains.

Classifiers trained on one domain's neurons perform poorly on others.

Hallucination mechanisms are domain-dependent, not universal.

Abstract

Recent work identifies a sparse set of "hallucination neurons" (H-neurons), less than 0.1% of feed-forward network neurons, that reliably predict when large language models will hallucinate. These neurons are identified on general-knowledge question answering and shown to generalize to new evaluation instances. We ask a natural follow-up question: do H-neurons generalize across knowledge domains? Using a systematic cross-domain transfer protocol across 6 domains (general QA, legal, financial, science, moral reasoning, and code vulnerability) and 5 open-weight models (3B to 8B parameters), we find they do not. Classifiers trained on one domain's H-neurons achieve AUROC 0.783 within-domain but only 0.563 when transferred to a different domain (delta = 0.220, p < 0.001), a degradation consistent across all models tested. Our results suggest that hallucination is not a single mechanism with a universal neural signature, but rather involves domain-specific neuron populations that differ depending on the knowledge type being queried. This finding has direct implications for the deployment of neuron-level hallucination detectors, which must be calibrated per domain rather than trained once and applied universally.