Citation Hallucinations

• Citation Hallucinations are errors where AI models generate fabricated or misattributed bibliographic references that lack real-world verification.
• They emerge from factors like over-generalization, training data redundancy, and retrieval gaps that lead to probabilistic assembly of citation details.
• Mitigation strategies such as retrieval-augmented generation, consistency-check frameworks, and post-hoc verification help ensure citation authenticity.

Citation hallucinations are errors that occur when generative artificial intelligence systems, particularly LLMs, produce citations or bibliographic references to sources that do not exist or misattribute content. These hallucinations manifest in diverse scholarly and professional contexts, including academic publishing, code comprehension, and legal analysis. Citation hallucinations undermine trust, obstruct verifiability, and may propagate misinformation when the presence of an apparently authoritative citation is itself fictitious or unsupported. The causes are rooted in probabilistic sequence modelling, gaps in retrieval architecture, and inadequate grounding of generated content in authentic sources.

1. Conceptual Foundations and Formal Definitions

Citation hallucinations are fundamentally distinct from general factual hallucinations in language modelling. In bibliographic contexts, a hallucinated citation is a reference to a non-existent work, often formatted correctly but lacking any real-world counterpart—composed by the stochastic assembly of title fragments, author names, journal identifiers, and other metadata without database verifiability (Glynn, 25 Mar 2025, &&&1&&&). In code comprehension and legal analysis, citation hallucinations further include misattributed links or references to supporting materials that do not substantively confirm the generated claim (Arafat, 13 Dec 2025, Hou et al., 2024).

Formally, given a generated segment $S_i$ with citation $R_i$, hallucination occurs when no entailment can be established: $\phi(R_i, S_i) = 0$ where $\phi$ is a natural language inference (NLI) verifier. The hallucination rate is then: $$\text{HallucinationRate} = 1 - \frac{1}{t} \sum_{i=1}^t \phi(\mathcal{F}(S_i), S_i)$$ If citations cannot be grounded in either external context or internal (parametric) model knowledge, they are classified as hallucinated (Shen et al., 21 Apr 2025).

2. Root Causes and Mechanisms

The emergence of citation hallucinations in generative LLMs is driven by several mechanisms:

• Over-generalization: LLMs are optimised to produce linguistically plausible outputs, not to verify factual existence. When prompted for citations, the model samples high-probability token sequences matching bibliographic patterns without database checking (Glynn, 25 Mar 2025).
• Training Data Redundancy: The frequency with which specific bibliographic records occur in the pretraining corpus correlates strongly with the model’s likelihood of correct recall. Highly cited items become verbatim “memorized” (reducing hallucinations), while items seen infrequently are synthesized probabilistically, increasing hallucination risk. Citation count acts as a measurable proxy for this redundancy (Niimi, 29 Oct 2025, Niimi, 12 Nov 2025).
• Memory Interference: When multiple highly cited records overlap in token-space (similar titles/authors), internal retrieval can produce hybrid or contaminated bibliographic outputs—mixing details of different real papers (Niimi, 12 Nov 2025).
• Retrieval and Architectural Gaps: In code and multi-document retrieval, hallucinations arise from failure to capture cross-file dependencies, sparse lexical matching, or lack of incorporating structural context—leading to citations of irrelevant or incomplete sources (Arafat, 13 Dec 2025).
• Prompt Leakage and Copy Artifacts: Interface labels (“Regenerate response”), present in AI outputs, can be inadvertently included in manuscripts, contributing to spurious citations (Glynn, 25 Mar 2025).
• Intrinsic and Extrinsic Gaps: In legal analyses, the taxonomy distinguishes internal incoherence or formatting errors (intrinsic) from mismatches in citation content (extrinsic hallucinations) (Hou et al., 2024).

3. Detection and Diagnostic Methodologies

State-of-the-art detection of citation hallucinations leverages a spectrum of techniques:

• Consistency-Check Frameworks: By querying the model about generated references (“Does this paper exist?”, “Who are the authors?”), detection algorithms exploit internal model representations to estimate groundedness. Ensemble methods blending direct and indirect consistency yield AUC up to 0.90 for hallucination discrimination—substantially better than naïve binary heuristics (Agrawal et al., 2023).
• Mechanistic Pathway Analysis: FACTUM decomposes transformer activations into attention updates, parametric-force scores, context alignment, and pathway alignment. Correct citations are characterised by strong parametric activation aligned with attentive evidence-synthesis, while hallucinations arise from discoordination or misalignment between these pathways. Detection performance improves by up to 37.5% AUC relative to baseline classifiers (Dassen et al., 9 Jan 2026).
• Faithfulness Metrics and Retrieval Evaluation: Both similarity-based (BERTScore, BARTScore) and entailment-based metrics (FactCC, SummaC, AutoAIS) are systematically benchmarked across three support levels: full, partial, and no support. Fine-grained metrics reveal that the “partial support” regime is most challenging, with current methods excelling at full-vs-none distinctions but struggling with partial hallucinations (Zhang et al., 2024, Zhang et al., 2024).
• Human-in-the-Loop and Chunk-Level Labeling: Manual annotation of supporting text spans and fine-grained gap categories (in legal tasks: claim hallucination, retrieval inaccuracy, citation omission) facilitates precise empirical error analysis, training of detectors, and statistical reporting (Li et al., 2023, Hou et al., 2024).

4. Mitigation Strategies and Defenses

Practical interventions for managing citation hallucinations span both system and policy levels:

• Full-Text Reference Deposit: Adoption of mandatory full-text deposit requirements enables reviewers and editors to verify the existence of all cited materials during manuscript submission. This protocol, inspired by TOP data standards and legal evidentiary practice, can reduce hallucination rates by up to 91.7% in early-adopter journals (Glynn, 25 Mar 2025).
• Retrieval-Augmented Generation (RAG): Conditioning LLM outputs on retrieved documents (or text spans) from reputable corpora (e.g., CrossRef, Wikipedia, source code repositories) grounds responses, reducing hallucination rates versus generation from ungrounded model parameters (Li et al., 2024, Arafat, 13 Dec 2025).
• Post-Hoc Citation Verification and Regeneration: Citation-Enhanced Generation (CEG) invokes iterative retrieval and NLI verification cycles for all model-generated claims, regenerating responses until each statement is backed by entailing citations (Li et al., 2024).
• Metadata-Aware Prompting and Abstention: Structured prompts requesting DOIs, URLs, or explicit abstentions when uncertain can discourage the model from fabricating references and encourage more conservative behavior in low-confidence scenarios (Niimi, 12 Nov 2025, Niimi, 29 Oct 2025).
• Hybrid Indexing & Parameter Attribution: Combining internal memorization checks with external retrieval databases as conditional sources for citation fields balances accuracy for highly cited (memorized) items with database-backed generation for rarer works (Niimi, 12 Nov 2025, Shen et al., 21 Apr 2025).
• Weighted Loss and Reference Calibration: Models fine-tuned with token-wise weighted objectives (INTRALIGN) for citation reliability, confidence calibration, and refusal when sources do not exist, demonstrate improved faithfulness and reduced hallucination and plagiarism rates (Shen et al., 21 Apr 2025).

5. Quantitative Characterisation and Evaluation Protocols

Empirical studies routinely employ multi-faceted metrics to quantify citation hallucination phenomena. Representative approaches include:

• Correlation and Regression Analysis: Linear and logistic regression over citation counts and factual consistency scores (using cosine similarity embeddings) establish redundancy scaling laws; high Pearson’s r ($r=0.75$, $p<.001$) evidences log-linear relationships between citation count and citation accuracy (Niimi, 29 Oct 2025, Niimi, 12 Nov 2025).
• ROC-AUC and Precision-Recall Metrics: Consistency filters and self-check frameworks for citation prediction reach ROC-AUC up to 0.90, precision rates above 80%, and balanced recall for legal and scientific domains (Agrawal et al., 2023, Hou et al., 2024, Dassen et al., 9 Jan 2026).
• Faithfulness and Extractiveness Scores: Automatically computed non-contradiction (ANLI) scores, BLEU/ROUGE metrics, and extractiveness profiles (coverage/density) quantitatively demonstrate that conditioning on cited text spans (CTS) reduces hallucination rates and improves match to gold-standard human citations (Li et al., 2023).
• Support-Level Classification and Retrieval Ranking: Tasks disentangle full, partial, and no-support, tabulating ROC-AUC and mean reciprocal rank (MRR) over annotated datasets (VeJudge, GenSearch) to expose failure modes neglected by binary classifiers (Zhang et al., 2024, Zhang et al., 2024).

6. Limitations, Open Challenges, and Future Directions

Despite notable progress, several limitations persist in citation hallucination research:

• Partial Support Sensitivity: Faithfulness metrics struggle to reliably discriminate partial support (PS) from none (NS) or full (FS); robust contrastive training with fine-grained annotations is recommended (Zhang et al., 2024, Zhang et al., 2024).
• Scale-Dependence of Mechanistic Signatures: Pathway alignment and attention signatures distinguishing correct from hallucinated citations evolve with model parameter count, requiring tailored detection for each architecture (Dassen et al., 9 Jan 2026).
• Annotation Overhead and Domain Adaptation: Human-in-the-loop CTS labeling remains labor-intensive; distant labeling and keyword-based retrieval, though practical, may miss semantically relevant but lexically divergent support (Li et al., 2023).
• Copyright Constraints in Full-Text Deposit: Implementation of full-text reference protocols depends on jurisdictional fair-use exemptions and may require standardized exception handling for closed-access content (Glynn, 25 Mar 2025).
• API Accessibility of Mechanistic Detectors: Advanced model-pathway metrics (FACTUM) require deep model introspection unavailable via most LLM APIs. Black-box proxy development is an open priority (Dassen et al., 9 Jan 2026).
• False Positives in Consistency Checks: Models may consistently (but incorrectly) reproduce plausible author names for hallucinated titles, limiting the ultimate fidelity of current self-consistency heuristics (Agrawal et al., 2023).

Continued research is anticipated in developing explainable, hybrid faithfulness metrics; extending domain-specific retrieval and annotation protocols; and refining transparent citation generation paradigms for both external and internal knowledge sources (Shen et al., 21 Apr 2025).

7. Policy, Practice, and Recommendations

Defensive measures against citation hallucinations are multifactorial. Key recommendations include:

• Mandate full-text reference deposit in academic publishing workflows, with upfront compliance verification (Glynn, 25 Mar 2025).
• Employ retrieval-augmented generation and post-hoc NLI verification in chatbot and RAG architectures (Li et al., 2024, Arafat, 13 Dec 2025).
• Leverage context-prior augmented generation tasks and INTRALIGN-style fine-tuning for explicit attribution and calibrated self-citation confidence (Shen et al., 21 Apr 2025).
• Integrate gap-detection taxonomies and fine-grained scoring for legal and medical domains (Hou et al., 2024).
• Build and use annotated training and evaluation resources capturing fine distinctions between partial and full citation support, supporting robust metric development (Zhang et al., 2024, Zhang et al., 2024).

Recognizing citation hallucinations as an inherent risk in probabilistic text generation, the field is advancing toward transparent, reliably verifiable citation workflows and mechanistic detection pipelines that safeguard the integrity of scholarly and professional output.