AI evidence extraction

Where Fake Citations Are Made: Tracing Field-Level Hallucination to Specific Neurons in LLMs

Yuefei Chen, Yihao Quan, Xiaodong Lin, Ruixiang Tang · ArXiv.org · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

2/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Multi-stage computational study combining: (1) empirical analysis of citation hallucination across 9 models and 108,000 generated references verified against OpenAlex; (2) probing analysis using linear logistic regression probes on hidden states to detect hallucination signals; (3) neuron-level localization using elastic-net regularization with stability selection on CETT values; (4) causal intervention via activation patching to test neuron suppression and enhancement effects..

Main result

The study finds that "author names fail far more often than other fields across all models and settings" and that "citation style has no measurable effect, while reasoning-oriented distillation degrades recall." Additionally, "probes trained on one field transfer at near-chance levels to the others, suggesting that hallucination signals do not generalize across fields." The researchers identify field-specific hallucination neurons and demonstrate that "amplifying these neurons increases hallucination, while suppressing them improves performance across fields."

Research paradigm

Empirical computational analysis with causal intervention experiments

Author conclusions

"Together, these findings show that citation hallucination is detectable from internal model signals, follows different patterns across bibliographic fields, and can be partly reduced through targeted neuron suppression without external retrieval." The authors further conclude that "citation hallucination arises from field-dependent internal representations and can be partly mitigated from within the model, without relying solely on external retrieval."

Risk of bias

Topic selection bias: 50 computer science topics may not represent other domains; Model selection bias: primary analysis on single model (Qwen2.5-32B-Instruct) with limited replication; Database coverage bias: OpenAlex may not index all valid citations, leading to false positive hallucination labels; Verification pipeline bias: GPT-5.4-mini used as grounded judge could introduce systematic errors; Topic-level selection bias: only 50 computer science topics used, may not represent hallucination patterns across other domains; Model selection bias: detailed analysis limited to Qwen2.5-32B-Instruct; cross-field transfer replicated only on one additional model (Mistral-Small-24B); Verification pipeline potential bias: two-stage verification using GPT-5.4-mini introduces dependency on another LLM's capabilities; Sample composition bias: intervention experiments conducted on randomly sampled topic subset, which authors acknowledge creates baseline discrepancy; Database coverage artifacts: OpenAlex indexing gaps may conflate genuine hallucinations with missing coverage; Small intervention sample size: n=5 references per condition limits statistical power for paired tests; Topic-level data leakage prevented by assigning all references from same topic to either training or test; Class imbalance addressed through downsampling to 1:1 positive-to-negative ratio in probe training; Potential generalization bias: single model architecture (Qwen2.5-32B-Instruct) used for primary probing and neuron analysis; Domain bias: evaluation limited to 50 computer science topics only

Limitations

"All intervention analyses are limited by the small number of structured fields (n=5), which constrains statistical power, and we do not evaluate whether neuron suppression affects output fluency beyond the JSON schema validity reported in Section 4.2
Additionally, probing and neuron localization use a single model (Qwen2.5-32B-Instruct) and 50 topics from computer science, leaving generalization to other architectures and domains untested."

Open questions raised

Generalization to other model architectures beyond Qwen and Mistral families
Applicability to non-computer science domains
Effects of neuron suppression on output fluency and naturalness beyond JSON schema validity
Scaling of FH-neuron interventions to production systems
Mechanisms underlying field-specific encoding in other structured generation tasks
Generalization of field-specific probing to other model architectures beyond Qwen2.5-32B-Instruct

Data: OpenAlex (public REST API): https://openalex.org/; 50 computer science research topics (listed in Appendix C); OpenAlex (used for reference verification via REST API); 108,000 generated citations across 9 models and 8 citation styles; 108,000 generated citations across 9 models verified against OpenAlex API (https://openalex.org)Code: Not mentioned in the paperExtracted from: pdfAgreement 59%

Explore related topics

Hallucination Control Research Integrity Empirical Study