Where Fake Citations Are Made: Tracing Field-Level Hallucination to Specific Neurons in LLMs
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866915947016617984 |
|---|---|
| author | Chen, Yuefei Quan, Yihao Lin, Xiaodong Tang, Ruixiang |
| author_facet | Chen, Yuefei Quan, Yihao Lin, Xiaodong Tang, Ruixiang |
| contents | LLMs frequently generate fictitious yet convincing citations, often expressing high confidence even when the underlying reference is wrong. We study this failure across 9 models and 108{,}000 generated references, and find that author names fail far more often than other fields across all models and settings. Citation style has no measurable effect, while reasoning-oriented distillation degrades recall. Probes trained on one field transfer at near-chance levels to the others, suggesting that hallucination signals do not generalize across fields. Building on this finding, we apply elastic-net regularization with stability selection to neuron-level CETT values of Qwen2.5-32B-Instruct and identify a sparse set of field-specific hallucination neurons (FH-neurons). Causal intervention further confirms their role: amplifying these neurons increases hallucination, while suppressing them improves performance across fields, with larger gains in some fields. These results suggest a lightweight approach to detecting and mitigating citation hallucination using internal model signals alone. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_18880 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Where Fake Citations Are Made: Tracing Field-Level Hallucination to Specific Neurons in LLMs Chen, Yuefei Quan, Yihao Lin, Xiaodong Tang, Ruixiang Computation and Language Artificial Intelligence LLMs frequently generate fictitious yet convincing citations, often expressing high confidence even when the underlying reference is wrong. We study this failure across 9 models and 108{,}000 generated references, and find that author names fail far more often than other fields across all models and settings. Citation style has no measurable effect, while reasoning-oriented distillation degrades recall. Probes trained on one field transfer at near-chance levels to the others, suggesting that hallucination signals do not generalize across fields. Building on this finding, we apply elastic-net regularization with stability selection to neuron-level CETT values of Qwen2.5-32B-Instruct and identify a sparse set of field-specific hallucination neurons (FH-neurons). Causal intervention further confirms their role: amplifying these neurons increases hallucination, while suppressing them improves performance across fields, with larger gains in some fields. These results suggest a lightweight approach to detecting and mitigating citation hallucination using internal model signals alone. |
| title | Where Fake Citations Are Made: Tracing Field-Level Hallucination to Specific Neurons in LLMs |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2604.18880 |