NeuroActiSep: Detecting Factual Hallucinations from Feed-Forward Neurons in a Single Pass

Ali Derogar Odolou; Reza Nazari; Mostafa Salehi

cs.CLSep 13, 2026

NeuroActiSep: Detecting Factual Hallucinations from Feed-Forward Neurons in a Single Pass

Authors: Ali Derogar Odolou, Reza Nazari, Mostafa Salehi

Organizations: School of Intelligent Systems Engineering College of Interdisciplinary Science and Technology University of Tehran, Tehran, Iran

Abstract

Hallucination in large language models reduces their reliability and slows adoption. Various white-box studies have used internal representations to detect patterns of truthfulness and factuality. A less-studied approach is to identify feed-forward neurons correlated with hallucination. We propose a method to rank feed-forward neurons at the final prompt token using a custom neuron selection dataset. We transfer the selected neuron identities to train hallucination classifiers on other factual question answering datasets. Our work provides empirical evidence that probes trained using the features from the selected neurons perform on par with probes trained on internal states. We also analyze the distribution of selected neurons and the effect of layer depth on detection performance.

Figures & tables

Fig. 1: We collect $u_{i,j}^{\ell}$ and $z_{i,j}^{\ell}$ to compute $\delta_{i,j}^{\ell}$ for ranking neurons as described in Section II-B

Model	Mistral	Qwen	Llama
Chosen 1%	0.878 $\pm$ 0.003	0.828 $\pm$ 0.001	0.845 $\pm$ 0.007
Final Layer MLP Output	0.858 $\pm$ 0.000	0.794 $\pm$ 0.003	0.837 $\pm$ 0.008
Final Layer Residual Stream	0.860 $\pm$ 0.003	0.804 $\pm$ 0.005	0.839 $\pm$ 0.001

TABLE I: AUROC Performance of Linear Probes on the Neuron Selection Dataset Using Three-Fold Cross Validation

Fig. 2: Distribution of neurons from the chosen set

\mathcal{S}_{1\%}

across layers and AUROC of linear NeuroActiSep on the neuron selection dataset for each depth cutoff

Dataset	Model	Truthful	Halluc.	Refusal	No Label
TruthfulQA	Mistral	532	254	29	2
	Qwen	501	308	8	0
	Llama	427	381	8	1
TriviaQA (Train)	Mistral	6818	2812	193	177
	Qwen	6528	3250	90	132
	Llama	7501	2230	90	179

TABLE II: Number of labels assigned to each dataset by our annotation pipeline

Method		TriviaQA			TruthfulQA
		Mistral	Qwen	Llama	Mistral	Qwen	Llama
Generation Scores	Perplexity	0.617	0.712	0.705	0.545 $\pm$ 0.037	0.522 $\pm$ 0.042	0.570 $\pm$ 0.036
	Mean Probability	0.615	0.697	0.693	0.537 $\pm$ 0.041	0.514 $\pm$ 0.048	0.563 $\pm$ 0.034
	Mean Logits	0.617	0.656	0.670	0.580 $\pm$ 0.036	0.570 $\pm$ 0.048	0.562 $\pm$ 0.042
	Min Logits	0.625	0.749	0.763	0.604 $\pm$ 0.028	0.588 $\pm$ 0.045	0.588 $\pm$ 0.046
P(True)		0.730	0.745	0.720	0.406 $\pm$ 0.026	0.510 $\pm$ 0.052	0.361 $\pm$ 0.017

TABLE III: AUROC performance of NeuroActiSep compared with baselines across datasets and models

Factual QA

See all 25 papers

Do Hallucination Neurons Generalize? Evidence from Cross-Domain Transfer in LLMs
Mar 27, 2026Snehit Vaddi, Pujith VaddiHallucination in Language Models LLM Hallucination Detection
Independent Researcher
Hallucination Neurons and Where to Find Them: An Investigation into the existence of Hallucination Neurons
Sep 24, 2026Huseyin Cavus, Sebin Sabu, Joshua Spear +2LLM Interpretability Hallucination in Language Models
Trakya University, Edirne, Türkiye · DRIVE, Great Ormond Street Hospital for Children NHS Foundation Trust, London, UK
Automatic Layer Selection for Hallucination Detection
May 25, 2026Xinpeng Wang, William X. Cao, Andrew Gordon Wilson +1Hallucination in Language Models LLM Hallucination Detection
University of Virginia · 2Walrus Security · 3New York University.

Mar 27, 2026J/K move · Enter open · S save

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

Snehit Vaddi, Pujith Vaddi

Independent Researcher

Hallucination in Language Models LLM Hallucination Detection LLM Interpretability Cross-Domain Generalization

Open paper AskPDF arXiv

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.

Abstract

Factual QA

Explore similar work

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

Hallucination Neurons and Where to Find Them: An Investigation into the existence of Hallucination Neurons

Automatic Layer Selection for Hallucination Detection