A literature review you can check citation by citation
Context generated this from open-access papers, unedited. Click a highlighted citation to see the source it matches and the passages extracted from it; the panel on the side is the verification panel from your workspace, read-only here.
Literature Review — 23 Sept 2026
Literature review · 1,163 words · 4 sources
Verify
Citations5 detectedClick any citation to verify
Citation Hallucination in Large Language Models: A Review of Prevalence, Mechanisms, and Mitigation
Introduction
The rapid incorporation of large language models (LLMs) into academic writing and scientific information retrieval has raised an urgent question: can the references these systems generate be trusted? Unlike conventional factual hallucination, citation fabrication threatens the specific infrastructure of scholarly verification—the chain linking a claim to a locatable, checkable source. This review synthesises four recent empirical studies that quantify, characterise, and attempt to mitigate this problem across different points in the research pipeline: model output auditing , automated detection architectures , retrospective forensic analysis of published surveys , and benchmarking of source attribution under uncertainty . Organised around four recurring concerns—the scale and stability of fabrication, its underlying mechanisms, the architectures proposed for detection, and the unresolved problem of epistemic calibration—the review establishes that citation hallucination is neither a marginal defect nor a uniformly understood phenomenon, but a structurally embedded and only partially tractable feature of current LLM-mediated scholarship.
The Scale and Persistence of Fabrication
A first point of convergence across the literature is that citation hallucination is common, measurable, and resistant to simple optimism about model improvement. Naser (2026) conducts the most extensive controlled audit to date, finding hallucination rates spanning 11.4% to 56.8% across ten commercial LLMs—a fivefold, statistically robust spread that undermines any assumption of a single "hallucination rate" for LLMs as a class. İlter (2026) approaches the problem from the opposite end of the pipeline, auditing 50 already-published AI survey papers rather than raw model output, and finds a 17.0% phantom citation rate that shows no significant temporal improvement over sixteen months (β1=+0.07 pp/month, p=0.72). Read together, these findings suggest that fabrication is not a transient artefact of early or under-trained models but a stable equilibrium that persists both at the point of generation and after human editorial mediation. This convergence is significant because it rules out a common assumption—that hallucination is simply a "growing pains" problem that newer model generations will resolve. Naser's finding that a newer model (Haiku-4.5) regressed relative to its predecessor reinforces İlter's equilibrium argument, indicating that scale and recency of release are not reliable proxies for citation reliability.
Mechanism: Distortion Rather Than Pure Invention
A second, more diagnostic theme concerns how hallucination occurs, and here the sources substantially agree despite differing units of analysis. Naser (2026) finds that 58.3% of hallucinated citations are "probable real" distortions of genuine papers, with outright fabrication accounting for only 6.0% of cases. İlter (2026) arrives at a strikingly compatible picture through independent forensic methods, showing that 78.5% of phantom citations in published surveys are "Syntax Errors"—correct titles paired with hallucinated identifying metadata (DOIs, page ranges)—while pure inventions ("Ghosts") constitute only 5.1%. İlter characterises this as a "lazy research assistant" pattern: models retrieve semantically correct content but fabricate the bureaucratic apparatus needed to locate it, severing what the author terms the "digital chain of custody." This mechanistic convergence across an experimental audit and an observational forensic study lends the distortion hypothesis considerable weight, though it complicates detection strategies premised on identifying wholly invented references, since the harder and more common problem is verifying near-correct citations rather than screening out fictitious ones. Tilwani et al. (2026) extends this picture into the attribution setting, showing that hallucination is overwhelmingly a factual error problem rather than a paraphrase-boundary issue (a 12.7:1 ratio of factual hallucinations to acceptable paraphrases), reinforcing that the core failure is precision of grounding rather than semantic incoherence.
Detection Architectures: From Heuristics to Multi-Agent Adjudication
Given this mechanistic profile, detection research has moved from binary verification toward richer, multi-source adjudication, though the sources diverge on how much infrastructure is required. Naser (2026) proposes lightweight, database-independent heuristics: multi-model consensus (≥3 models agreeing) achieves 95.6% accuracy, and a GBM classifier trained purely on 27 bibliographic string features reaches an AUC of 0.876, requiring no external API calls. In contrast, Li et al. (2026) argue that binary found/not-found detection is inherently brittle and propose CITETRACER, a cascading multi-agent system that retrieves evidence across multiple channels and routes ambiguous cases to class-specialist judgers operating over a 12-code taxonomy distinguishing REAL, POTENTIAL, and HALLUCINATED citations. This architecture achieves 97.1% accuracy on a synthetic benchmark and detects 97.1% of fabrications in a real-world set drawn from ICLR 2026 submissions. The tension between these approaches is instructive rather than contradictory: Naser's heuristics prioritise deployability and speed, useful as pre-screening filters, while Li et al.'s system prioritises diagnostic granularity for cases—such as İlter's "Broken Link" category—where a citation is neither clearly valid nor clearly fabricated. Neither study, however, tests its method against the other's benchmark, leaving open whether lightweight heuristics could match multi-agent adjudication on real-world, non-synthetic fabrications, or whether the POTENTIAL class that CITETRACER handles least reliably (F1 of 95.8) corresponds precisely to the "Sloppy" and "Unknown" categories that İlter finds hardest to resolve.
The Abstention–Hallucination Tradeoff
A distinct and less resolved theme concerns not whether models fabricate, but whether they know when not to answer. Tilwani et al. (2026) reframes the problem around epistemic safety, introducing a joint Abstention Rate/Hallucination Rate metric and demonstrating a troubling tradeoff: retrieval-augmented approaches that reduce hallucination (Advanced RAG lowers HR from 87.6% to 65.4% relative to Naive RAG) do so by eliminating abstention almost entirely, converting cautious refusal into confident error. This finding complicates the optimistic reading of retrieval-augmented mitigation implicit in Naser's and Li et al.'s detection-oriented frameworks, since improving retrieval does not straightforwardly improve reliability if it also suppresses appropriate uncertainty signalling. Tilwani et al. further show that hallucination is structurally elevated in specialised domains (Quantum Computing, Biomolecules) even after controlling for data volume, suggesting an epistemic equity concern absent from the other three studies, which largely treat domain as a control variable rather than a locus of systematic risk.
Conclusion
Across these four studies, a coherent empirical picture emerges: citation hallucination is prompt-sensitive and highly variable across models , persistent at a stable structural rate in published output , predominantly a matter of metadata distortion rather than wholesale invention , and entangled with a poorly calibrated abstention mechanism that current mitigation strategies may inadvertently worsen . Detection research has correspondingly diversified from simple existence-checking toward multi-source, taxonomy-aware adjudication , yet no study directly compares lightweight classifier-based screening against multi-agent verification on a shared, real-world benchmark, leaving the relative cost-effectiveness of these approaches unresolved. More fundamentally, the field lacks a unified account of why hallucination varies so sharply across models and domains—Naser explicitly treats this as correlational rather than causal, and Tilwani et al.'s domain-specialisation finding remains observational. The interaction between retrieval augmentation, abstention calibration, and long-term citation-graph integrity—captured provocatively but only speculatively in İlter's Muller's-Ratchet decay model—represents the most significant open problem: current evidence establishes that fabrication is common and mechanistically patterned, but not yet how to prevent it without trading one failure mode for another.
Verification Workflow
0 verified0 questionable0 needs revision
Source Trust
Exclude weak sources from future regenerations.
Li et al., 2026
Tilwani et al., 2026
İlter, 2026
Naser, 2026
Section Review
Outputs are drafts. You still do the thinking — this page only shows what there is to check.
Papers used are published under CC BY 4.0: Naser, M.Z. (2026); İlter, H.K. (2026); Tilwani, D. et al. (2026); Li, M. et al. (2026)