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

Snehit Vaddi; Pujith Vaddi

cs.CLMar 27, 2026

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

Authors: Snehit Vaddi, Pujith Vaddi

Organizations: Independent Researcher

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.

Figures & tables

Figure 1: Overview of the cross-domain H-neuron transfer protocol. For each model, we extract CETT-based neuron activation features across 6 domains, train per-domain L1-regularized logistic regression classifiers, and evaluate every classifier on every domain to construct a

6\times 6

transfer matrix.

Model	Gen.	Leg.	Fin.	Sci.	Mor.	Code
Qwen-3B	.896	.795	.882	.662	.836	.597
Nemotron-4B	.876	.807	.677	.753	.872	.563
Phi-3.5	.902	.768	.838	.871	.899	.557
Mistral-7B	.824	.862	.767	.929	.818	.666
Llama-8B	.805	.747	.890	.797	.805	.542
Mean	.861	.796	.811	.802	.846	.585

Table 1: Within-domain AUROC for H-neuron classifiers (direct prompting). Each value represents a classifier trained and tested on the same domain (held-out test set). All values except code substantially exceed chance (0.5). Bootstrap point estimates with CIs are reported in Appendix D .

Figure 2: Cross-domain transfer matrix (mean AUROC across 5 models, direct prompting). Diagonal entries (within-domain) are consistently higher than off-diagonal entries (cross-domain). The transfer gap $\Delta=0.220$ is statistically significant ( $p<0.001$ , permutation test).

Train $\downarrow$ Test $\to$	Gen.	Leg.	Fin.	Sci.	Mor.	Code
General	.861	.582	.617	.604	.634	.510
Legal	.525	.796	.556	.698	.552	.517
Financial	.526	.595	.811	.560	.602	.494
Science	.559	.722	.542	.802	.572	.523
Moral	.601	.578	.618	.547	.846	.530
Code	.470	.507	.508	.542	.508	.585

Table 2: Mean cross-domain AUROC across 5 models (direct prompting). Shaded diagonal entries are within-domain; off-diagonal entries are cross-domain transfers. The diagonal consistently dominates.

Figure 3: Within-domain (dark) vs. cross-domain (light) mean AUROC by model. The gap is consistent across all models and architectures.

Direct prompting
Domain	Qwen	Nem.	Phi	Mis.	Llama
General	.896	.876	.902	.824	.805
Legal	.795	.807	.768	.862	.747
Financial	.882	.677	.838	.767	.890
Science	.662	.753	.871	.929	.797
Moral	.836	.872	.899	.818	.805

Table 3: Within-domain AUROC under direct vs. CoT prompting.

\dagger

indicates pairs where the model produced identical responses under both strategies, yielding identical classifier results.

Figure 4: Comparison of within-domain AUROC under direct (left) and CoT (center) prompting, with the difference shown on the right. CoT effects are heterogeneous across models and domains.

Figure 5: Domain specificity (transfer gap

\Delta

) by model size. Domain specificity does not consistently increase or decrease with model scale in the 3B-8B range.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Domain	Dataset	Split	Samples Used	Task Type	Answer Format
General QA	TriviaQA	test	1000	Open-ended factual QA	Free-form text
Legal	CaseHOLD	test	1000	Holding identification	5-way multiple choice
Financial	FinancialPhraseBank	train	1000	Sentiment classification	3-way classification
Science	ARC-Challenge	test	1000	Science reasoning	4-way multiple choice
Moral	ETHICS (commonsense)	test	1000	Moral judgment	Binary classification
Code	Devign	test	1000	Vulnerability detection	Binary classification

Appendix

Table 4: Dataset details for each domain. Samples Used refers to questions presented to each model; see Table 10 for evaluation counts after filtering.

Train $\downarrow$ / Test $\to$	General	Legal	Financial	Science	Moral	Code
General	.805	.571	.690	.563	.876	.529
Legal	.607	.747	.707	.734	.604	.588
Financial	.593	.659	.890	.606	.763	.534
Science	.593	.754	.519	.797	.574	.520
Moral	.602	.561	.641	.522	.805	.498
Code	.489	.539	.558	.463	.651	.542

Appendix

Table 5: Transfer matrix for Llama-3.1-8B-Instruct (direct prompting). Within-domain diagonal in bold.

Train $\downarrow$ / Test $\to$	General	Legal	Financial	Science	Moral	Code
General	.824	.576	.685	.742	.725	.481
Legal	.351	.862	.607	.873	.374	.448
Financial	.642	.579	.767	.731	.586	.514
Science	.393	.845	.627	.929	.513	.475
Moral	.651	.631	.559	.384	.818	.527
Code	.411	.436	.423	.593	.529	.666

Appendix

Table 6: Transfer matrix for Mistral-7B-Instruct-v0.3 (direct prompting).

Train $\downarrow$ / Test $\to$	General	Legal	Financial	Science	Moral	Code
General	.876	.536	.618	.552	.444	.586
Legal	.579	.807	.406	.582	.779	.516
Financial	.498	.535	.677	.391	.737	.498
Science	.592	.738	.590	.753	.718	.647
Moral	.524	.563	.663	.669	.872	.512
Code	.488	.644	.513	.470	.210	.563

Appendix

Table 7: Transfer matrix for Nemotron-Mini-4B-Instruct (direct prompting).

Train $\downarrow$ / Test $\to$	General	Legal	Financial	Science	Moral	Code
General	.902	.597	.451	.665	.511	.474
Legal	.467	.768	.491	.642	.256	.548
Financial	.516	.645	.838	.532	.461	.404
Science	.604	.605	.374	.871	.332	.423
Moral	.605	.490	.558	.625	.899	.588
Code	.462	.465	.449	.631	.488	.557

Appendix

Table 8: Transfer matrix for Phi-3.5-mini-instruct (direct prompting).

Train $\downarrow$ / Test $\to$	General	Legal	Financial	Science	Moral	Code
General	.896	.633	.641	.498	.614	.479
Legal	.619	.795	.567	.660	.747	.485
Financial	.383	.555	.882	.541	.460	.519
Science	.614	.668	.598	.662	.724	.548
Moral	.624	.642	.667	.534	.836	.524
Code	.500	.450	.597	.552	.662	.597

Appendix

Table 9: Transfer matrix for Qwen2.5-3B-Instruct (direct prompting).

Domain	Qwen	Nem.	Phi	Mis.	Llama
General	160	200	212	142	90
Legal	230	104	180	82	88
Financial	128	86	110	112	80
Science	100	106	56	76	64
Moral	124	66	192	90	26
Code	264	240	256	103	182

Appendix

Table 10: Evaluation sample sizes per model and domain.

Model	Domain	AUROC [95% CI]	CV (mean $\pm$ std)	Perm. $p$	Verdict
Qwen2.5-3B	General	.925 [.882, .963]	.920 $\pm$ .013	.000	ROBUST
	Legal	.788 [.725, .843]	.794 $\pm$ .035	.000	ROBUST
	Financial	.749 [.655, .839]	.806 $\pm$ .026	.000	ROBUST
	Science	.783 [.691, .866]	.763 $\pm$ .076	.000	ROBUST
	Moral	.818 [.746, .889]	.811 $\pm$ .020	.000	ROBUST
	Code	.604 [.538, .673]	.621 $\pm$ .040	.000	ROBUST

Appendix

Table 11: Robustness analysis for all experiments under direct prompting. AUROC is the bootstrap point estimate with 95% CI (1,000 resamples). CV reports 5-fold cross-validated AUROC.

Model	Domain	AUROC [95% CI]	CV (mean $\pm$ std)	Perm. $p$	Verdict
Qwen2.5-3B	General	.924 [.880, .961]	.942 $\pm$ .016	.000	ROBUST
	Legal	.793 [.738, .850]	.788 $\pm$ .017	.000	ROBUST
	Financial	.830 [.750, .901]	.809 $\pm$ .009	.000	ROBUST
	Science	.607 [.499, .724]	.692 $\pm$ .055	.000	WEAK
	Moral	.850 [.779, .908]	.836 $\pm$ .021	.000	ROBUST
	Code	.646 [.574, .713]	.628 $\pm$ .041	.050	WEAK

Appendix

Table 12: Robustness analysis for all experiments under CoT prompting.

\dagger

marks experiments identical to direct prompting (response caching). See Table 11 caption for details.

Hallucination in Language Models

Hallucination in language models studies how systems produce unsupported or false claims and how to detect, explain, and reduce them.

See all 247 papers

NeuroActiSep: Detecting Factual Hallucinations from Feed-Forward Neurons in a Single Pass
Sep 13, 2026Ali Derogar Odolou, Reza Nazari, Mostafa SalehiFactual QA LLM Hallucination Detection
School of Intelligent Systems Engineering College of Interdisciplinary Science and Technology University of Tehran, Tehran, Iran
Domain-Specific Hallucination Detection in Large Language Models
Sep 10, 2026Varun Teja Chundru, Debasmita BiswasLanguage Model Error Detection LLM Hallucination Mitigation
Department of Computer Science Purdue University Fort Wayne
CrossHallu: Do Hallucination Signals Generalize Across Languages and Domains in Large Language Model's Internals?
Jul 4, 2026Aisha Alansari, Malak Alkhorasani, Hamzah LuqmanMultilingual Language Model Evaluation LLM Hallucination Detection
King Fahd University of Petroleum and Minerals · Imam Abdulrahman bin Faisal University

Sep 13, 2026J/K move · Enter open · S save

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

Ali Derogar Odolou, Reza Nazari, Mostafa Salehi

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

Factual QA LLM Hallucination Detection Hallucination in Language Models LLM Interpretability Neuron-Level Interpretability

Open paper AskPDF arXiv

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.

Abstract

Hallucination in Language Models

Explore similar work

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

Domain-Specific Hallucination Detection in Large Language Models

CrossHallu: Do Hallucination Signals Generalize Across Languages and Domains in Large Language Model's Internals?