cs.CLOct 21, 2026
SaveCiteAnchor: Source-Aware Grounding for Hallucinated Citation Detection and Resolution
Organizations: University of Wisconsin–Madison · University of Waterloo · Texas A&M University
Abstract
Hallucinated citations make it harder to trace sources and verify the evidence behind scientific claims. Detecting them is difficult because bibliographic records can be incomplete or differ across source types and publication versions. A plausible match alone neither validates the citation nor establishes its underlying source. We introduce CiteResolve, a benchmark of 423 real-world hallucinated citations annotated for source recoverability. We further propose CiteAnchor, a source-aware grounding framework that adapts retrieval and field-level verification to each source type and separately determines citation validity and whether a unique source can be recovered. On an existing detection benchmark, CiteAnchor achieves 0.841 F1 with GPT-5.6 Luna and 0.784 with an open-weight backbone, outperforming the baseline with the same backbones by 7.9 and 36.8 points. On CiteResolve, it achieves 74.7% and 70.7% end-to-end task success, outperforming the corresponding web search baselines by 16.3 and 65.3 percentage points. Our code and data are available at https://github.com/oldwizard7/CiteANCHOR_code.
Figures & tables
| Cited as | Recoverable | Unrecoverable | Total |
|---|---|---|---|
| Paper | 225 | 163 | 388 |
| Book | 3 | 1 | 4 |
| Web resource † | 24 | 7 | 31 |
| Total | 252 | 171 | 423 |
Table 1: CiteResolve at a glance. † Web pages, datasets, software, and technical reports. Publication details refer to venue, year, volume, pages, or link. Webpage details involve the page title, the author or organization, and the address.
Figure 1: Overview of CiteAnchor with a running example. CiteAnchor retrieves source-specific evidence, verifies citation fields against the retrieved records, and resolves the underlying source when one work is uniquely supported. Otherwise, it abstains from source resolution.
| Confusion matrix | Metrics | |||||||
| Method | TP | FN | FP | TN | Acc. | Prec. | Rec. | F1 |
| Published results for reference ( Shi et al., 2026 ) | ||||||||
| GPTZero # Adam et al. (2026) | 338 | 129 | 1,358 | 1,531 | 0.557 | 0.199 | 0.724 | 0.313 |
| Gemini-3-Pro # Google DeepMind (2026) | 351 | 116 | 412 | 2,477 | 0.843 | 0.460 | 0.752 | 0.571 |
| GPT-5.2 # OpenAI (2025) | 366 | 101 | 1,379 | 1,510 | 0.559 | 0.210 | 0.784 | 0.331 |
| Claude-Sonnet-4.5 # Anthropic (2025) | 349 | 118 | 756 | 2,133 | 0.740 | 0.316 | 0.747 | 0.444 |
Table 2: Detection on the CiteAudit real-world test set. Published results ( # ) from Shi et al. (2026) are included for reference. CiteAudit’s published result uses Gemini-3-Flash ( Google DeepMind, 2025 ) . The lower blocks compare methods using the indicated backbone. Bold marks the best metric within each backbone block.
| Recoverable (252) | Unrecoverable (171) | Overall | ||
| Method | Identified | Recovered | Identified | Task success |
| Backbone: GPT-5.6 Luna ( OpenAI, 2026a ) | ||||
| Direct Prompting | 23.0 | 36.2 | 19.3 | 12.8 |
| Web Search | 85.3 | 90.2 | 31.0 | 58.4 |
| CiteAnchor | 65.9 | 95.2 | 92.4 | 74.7 |
| Backbone: OpenResearcher-30B-A3B ( Li et al., 2026b ) | ||||
Table 3: Source resolution on CiteResolve under the end-to-end protocol in Section 3.3 . All values are percentages. Metrics are detailed in Section 5.1 . Bold marks the best value within each backbone block.
Figure 2: Commercial search use and estimated API cost over the 3,356 CiteAudit citations. (a) Number of SerpApi searches, including additional searches during source resolution. (b) Estimated combined cost of LLM calls and SerpApi searches using public API pricing. Hatching indicates search costs.
| Variant | CiteAudit F1 | CiteResolve Success |
|---|---|---|
| w/o Chain | 0.643 | 59.3 |
| w/o Search Engines | 0.734 | 72.6 |
| w/o Structured APIs | 0.773 | 53.2 |
| w/o Type Adaptation | 0.761 | 66.9 |
| CiteAnchor | 0.841 | 74.7 |
Table 4: Ablations with GPT-5.6 Luna. Resolution success is reported as a percentage. w/o Chain replaces staged verification and resolution with one LLM judgment over the same records. w/o Search Engines disables Google and Google Scholar searches. w/o Structured APIs removes bibliographic APIs while retaining search engines and cited webpages. w/o Type Adaptation uses generic retrieval and verification rules for all source types.
Figure 3: Two CiteResolve cases where CiteAudit, Web Search, and Direct Prompting fail with GPT-5.6 Luna. Red text marks incorrect citation fields, and red boxes show excerpts from the baselines’ explanations. (a) A real title is paired with an incorrect author list. CiteAnchor detects the contradiction and recovers the source. (b) An unrecoverable citation is repeated on the web without a supporting bibliographic record. CiteAnchor rejects the mismatched identifier and returns None .
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Additional statistics of CiteResolve . Left: origin of the gold record for the 252 Recoverable citations. “Cited web page” denotes cases where the cited webpage itself serves as the gold source; “Other” groups the remaining record sources. Middle: research areas of all 423 citations. Right: type of source cited by all 423 citations.
| Recoverable | Unrecoverable | |
|---|---|---|
| Contains identifier | 124 (49%) | 59 (35%) |
| Valid arXiv ID | 73 of 84 (87%) | 35 of 39 (90%) |
| Title match | 72 of 73 (99%) | 0 of 35 (0%) |
Table 5: Identifier evidence in CiteResolve . The first row reports the share of citations containing an arXiv ID, DOI, or URL. The next two rows focus on citations containing an arXiv ID.
| Bibliographic cue | Type |
|---|---|
| Academic identifier, preprint host, venue, volume/pages, or author–venue pattern | paper |
| ISBN, publisher, edition, handbook, encyclopedia, or chapter information | book |
| Technical/working paper, thesis, standard, or institutional report | report |
| Repository host such as GitHub, GitLab, Bitbucket, or Hugging Face model | repository |
| Dataset host such as Kaggle, Zenodo, figshare, Dryad, or OpenML | dataset |
| General webpage, blog, documentation, news, or accessed-date pattern | web page |
Table 6: Deterministic cues used for source-type routing. Cues are applied in a fixed priority order; the LLM backbone classifies only citations not resolved by these rules.
| Source type | Primary retrieval | Fallback retrieval | Evidence passed to verification |
|---|---|---|---|
| Paper | Identifier lookup and structured scholarly sources, including Crossref, OpenAlex, Semantic Scholar, arXiv for preprints, and Europe PMC for biomedical citations | Google Scholar, then Google | Bibliographic records with author metadata; generic web hits and reference-list-only Scholar entries are excluded |
| Book | Identifier lookup and scholarly sources, augmented with Google Books | Google Scholar, then Google | Book records whose title and author evidence are sufficient for identity comparison; differences across editions are allowed |
| Web resource | The cited address, or an archived copy when unavailable, together with identifier lookup | Google, then Google Scholar | The cited page or records tied to that address whose title provides evidence about the cited resource |
Table 7: Source-aware retrieval routes. Source type determines the primary evidence sources and the order in which retrieval is expanded. All routes first consult previously retrieved records and the offline ACL Anthology. Retrieved records are passed to evidence verification, which determines whether they support the cited work.
| Field | Comparison | match | missing | conflict |
|---|---|---|---|---|
| Title | Normalized word overlap and containment | The titles agree after normalization, or one is a clear shortened or truncated form of the other | The titles overlap, but not enough to establish either agreement or contradiction | The titles are clearly different |
| Authors | Parsed surname and given-name agreement; name order and “et al.” are ignored | The cited names substantially agree with the record | The author evidence is partial, incomplete, or otherwise inconclusive | A reliable author list contradicts the cited authors |
| Venue | Normalized venue names, including common aliases and abbreviations | The venues are equivalent | The venue is absent or the comparison is inconclusive | The venues are clearly different |
| Year | Publication year, allowing legitimate differences across versions | The years are compatible with the same work | The year is unavailable | The years are inconsistent with the same work |
| Identifier | Resolved DOI, arXiv identifier, or URL | The identifier points to the same work | The identifier is absent, incomplete, or insufficient to establish identity | The identifier resolves to a different work |
Table 8: Field-level evidence used in citation verification. missing denotes insufficient evidence to establish either a match or a conflict ; it does not require the field itself to be absent. These field outcomes are combined with source type and publication-version evidence when making the final validity decision.
| Decision pattern | Decided by | Citations |
| Real : the citation is supported | ||
| Clear supporting evidence | rules | 2,378 |
| Supporting evidence requiring judgment | LLM | 546 |
| Fake : the citation contains contradicted fields | ||
| Clear identity contradiction | rules | 141 |
| Identity contradiction requiring judgment | LLM | 71 |
Table 9: Decision patterns in evidence verification on the 3,356 CiteAudit citations with GPT-5.6 Luna. Deterministic rules make the final decision for 2,692 citations, while LLM judgment resolves the remaining 664.
More in this topic
Citation Verification
Citation verification checks whether cited sources exist, are accurately represented, and actually support the claims attached to them.
See all 49 papers