CiteAnchor: Source-Aware Grounding for Hallucinated Citation Detection and Resolution

1 Introduction

• Incomplete records and differences across source types and publication versions make simple matching unreliable; a plausible match may conceal an erroneous field or fail to establish a unique source.A valid citation may omit a database field, while an invalid citation may match a record that omits the erroneous field.Source: §1 · pp. 1–2
• CiteResolve evaluates whether real-world hallucinated citations have a uniquely supported source, with verified corrections for recoverable cases and appropriate abstention otherwise.The benchmark evaluates detection and source resolution together.Source: §1 · p. 2
• CiteAnchor uses type-specific retrieval, deterministic bibliographic checks, and LLM reasoning, checking fields before separately deciding citation validity and source identity.It shares retrieved evidence across discovery, verification, and resolution, with a targeted follow-up search when an error has no supported source.Source: §1 · p. 2
• CiteAnchor improves detection F1 over same-backbone CiteAudit by 7.9 and 36.8 points, and CiteResolve success over Web Search by 16.3 and 65.3 percentage points.The results cover proprietary and open-weight backbones.Source: §1 · p. 2

2 Related Work

• CiteAudit, CiteTracer, CiteCheck, and HALLMARK address bibliographic verification using retrieval, field checks, or evaluation of metadata errors and false positives.CiteTracer distinguishes missing information from contradictions; CiteCheck evaluates labels on synthetic corruptions.Source: §2 · p. 2
• CiteAnchor builds on retrieval and verification strategies while extending evaluation to recovered-source correctness and justified abstention.The preceding related work discusses verification and source resolution as distinct evaluation concerns.Source: §2 · pp. 2–3
• Citation resolution has been studied for noisy reference matching, context-based attribution, and citation support, but these settings differ from checking a bibliography entry alone.CiteME and CiteGuard start from surrounding text; ALCE and CiteFix evaluate citation support or correction against retrieved documents.Source: §2 · p. 3

3 The CITERESOLVE Benchmark

• CiteResolve pairs real-world hallucinated citations with verified source-resolution targets, requiring correct detection and a correct recovery or abstention for success.The benchmark contains 423 citations: 252 Recoverable and 171 Unrecoverable.Source: §3 · p. 3, §3.1 · p. 3

3.1 Benchmark Construction

• The benchmark retains CiteAudit’s real test set and Fake labels, then manually reviews and filters Fake cases for source-resolution annotation.Recoverable cases receive a verified corrected citation; Unrecoverable cases receive None.Source: §3.1 · p. 3

3.2 Annotation Criteria

• For papers, near-exact titles are the main recovery basis; close-but-not-exact titles also require an identifier, and author or year differences alone need not prevent recovery.Incorrect fields still need correction in the returned citation.Source: §3.2 · p. 3
• Book recovery requires title and author agreement, while web-resource recovery follows the URL and compares available page title and author or organization information.A missing identifiable author alone does not rule out web-resource recovery; an explicit author mismatch remains a discrepancy.Source: §3.2 · pp. 3–4
• None means no candidate meets the criteria or multiple distinct works remain plausible, not that no related work exists.The evaluation supplies only the citation string and scores resolution on the annotated subset, penalizing missed Fake detections.Source: §3.2 · p. 4, §3.3 · p. 4

4 The CITEANCHOR Framework

• CITEANCHOR connects source discovery, evidence verification, and source resolution, sharing retrieved records while combining deterministic checks with LLM reasoning.Field comparisons are separated from decisions about citation validity and source identity.Source: §4 · p. 4

4.1 Source Discovery

• Source discovery extracts available citation fields, identifies whether the source is a paper, book, or web resource, and retrieves records using relevant tools.Retrieval can revise a query when incorrect fields obscure a match.Source: §4.1 · p. 4

4.2 Evidence Verification

• Evidence verification compares cited fields with available records, distinguishing agreement, contradiction, and missing evidence while accounting for source type and legitimate publication-version differences.Deterministic checks handle clear cases; the LLM assesses unresolved comparisons.Source: §4.2 · p. 4

4.3 Source Resolution

• Source resolution groups records of the same work and proposes a correction only when one candidate remains supported; otherwise, it returns None.When no candidate is supported, the module adds one targeted web search before checking the evidence again.Source: §4.3 · p. 5

5.1 Experimental Setup

• The evaluation uses CiteAudit for detection and CITERESOLVE for end-to-end resolution, comparing two backbones against direct-prompting and CiteAudit or Web Search baselines.Resolution task success counts correct detection followed by correct recovery or abstention across all 423 citations.Source: §5.1 · pp. 5–6

5.2 Citation Detection on the CiteAudit Real-World Test Set

• 0.841 F1 with GPT-5.6 Luna improves on CiteAudit’s 0.762, while 0.784 with OpenResearcher-30B-A3B improves on CiteAudit’s 0.416.The reported comparison uses the same backbone for each method.Source: Table 2 · p. 6, §5.2 · p. 6

5.3 Source Resolution on CITERESOLVE

• Under the end-to-end protocol, overall task success was 74.7%, exceeding Web Search by 16.3 percentage points; with OpenResearcher-30B-A3B, CITEANCHOR reached 70.7%, versus 5.4% for Web Search and 3.3% for Direct Prompting.Source: Table 3 · p. 7, §5.3 · p. 7

5.2 Citation Detection on the CiteAudit Real-World Test Set

• 67.6% and 73.8% fewer SerpApi calls with GPT-5.6 Luna and OpenResearcher-30B-A3B, respectively, were used across 3,356 citations.With GPT-5.6 Luna, estimated combined LLM and commercial-search cost fell from approximately $36 to $12 per run.Source: Figure 2 · p. 7, §5.2 · p. 7

6.1 Ablation Studies

• Detection F1 falls from 0.841 to 0.643 and resolution success from 74.7% to 59.3% without staged verification and separate validity and identity decisions over the same records.This comparison evaluates the full decision procedure, including field checks and instructions.Source: §6.1 · p. 7
• Removing search engines lowers detection F1 from 0.841 to 0.734, while removing structured APIs lowers resolution success to 53.2% and detection F1 to 0.773.The different loss patterns suggest broad web search helps detect errors, while structured bibliographic records are especially valuable for identifying recoverable sources.Source: §6.1 · pp. 7–8, Table 4 · p. 8
• Generic rather than source-type-specific retrieval and verification reduces detection F1 to 0.761 and resolution success to 66.9%.The authors attribute the benefit to adapting both evidence collection and interpretation to the source type.Source: §6.1 · p. 8

6.2 Case Studies

• With a real title but an incorrect author list, CiteAnchor detects the contradiction and recovers the source, while all three baselines accept the citation.Seven of the ten cited authors are absent from the paper’s complete 13-author list; the matching title supports recovery.Source: §6.2 · p. 8, Figure 3 · p. 8
• When web pages repeat an unrecoverable citation but bibliographic records do not establish a match, CiteAnchor rejects the mismatched identifier and returns Fake with None.The identifier resolves to a different paper, so repetition on a webpage does not establish that the cited work exists.Source: §6.2 · p. 8, Figure 3 · p. 8

• CiteResolve benchmarks source resolution for hallucinated citations, while CiteAnchor separates citation validity from source identity.Source: §7 · p. 8
• Across proprietary and open-weight backbones, CiteAnchor improves hallucinated-citation detection and end-to-end source resolution while reducing commercial search use.Source: §7 · p. 8
• The authors conclude that reliable verification should recover a source when uniquely supported and otherwise abstain from correction, giving reviewers retrieved evidence to inspect.Source: §7 · p. 8

• CiteResolve is dominated by paper citations, so its results provide stronger evidence for papers than for books or web resources.Source: Limitations · p. 9
• Broader coverage across disciplines and languages would help assess how well the framework transfers to other settings.Source: Limitations · p. 9
• The task evaluates bibliographic validity and source identity, not whether the cited work supports the surrounding claim.A correct reference may still appear in an unsupported context.Source: Limitations · p. 9
• Resolution reflects records and services available during the experiments, and Unrecoverable means no unique source is supported under the authors’ criteria, not that no related work exists.Source: Limitations · p. 9

Ethical Considerations

• CiteAnchor is intended to help reviewers detect and examine problematic references, not to replace authors’ responsibility to verify them.A Fake prediction concerns the citation itself and does not establish LLM use, deliberate fabrication, or misconduct.Source: Ethical Considerations · p. 9
• A locally hosted backbone can keep model inference local, but external searches still transmit queries to service providers.Deployment on confidential submissions requires model and retrieval services that meet venue confidentiality requirements.Source: Ethical Considerations · p. 9

A.1 Construction

• CiteResolve retains citations labeled Fake in CiteAudit’s real-world test set and manually annotates whether each source is uniquely identifiable, yielding 423 citations after quality control.Source: §A.1 · p. 11
• Annotators begin with pooled top-three records from Crossref, OpenAlex, and Semantic Scholar, then may broaden searches or resolve identifiers when no record matches.CiteAnchor was not used during annotation; a second annotator independently re-annotated the citations, with disagreements resolved by discussion.Source: §A.1 · p. 11
• Recoverability requires evidence that consistently identifies one source; for papers, title evidence carries the most weight, while an author list alone is insufficient.Identifiers are resolved and checked against the remaining citation fields; multiple plausible sources mean Unrecoverable.Source: §A.1 · p. 11
• Different versions of the same work count as one source when identity is clear, and minor differences or missing fields alone do not make a citation unrecoverable.Source: §A.1 · p. 11
• For each Recoverable citation, the verified title, authors, venue, year, and available identifier form the gold target for evaluation.Source: §A.1 · p. 11
• For web resources lacking standard bibliographic metadata, the gold source is the resolved webpage, dataset page, or repository, with unavailable fields left empty.Source: §A.1 · p. 11

Composition

• Of 423 citations, 388 refer to papers, 31 to web resources, and 4 to books.The web resources comprise 17 webpages, 6 datasets, 6 reports, and 2 software repositories.Source: §A.2 · p. 12

Benchmark coverage

• The benchmark spans multiple research areas, with computing the largest, and includes journal or conference papers, preprints, and other source types.Gold records for Recoverable citations most often come from OpenAlex, Crossref, or arXiv.Source: §A.2 · p. 12

Identifiers

• Among Unrecoverable citations, 35 of 39 arXiv IDs are valid, yet none of the resolved papers matches the cited title.Identifiers must be resolved and checked against the remaining citation evidence.Source: §A.2 · p. 12

B.1 Common Setup

• Each backbone comparison uses the same backbone, and CITEANCHOR applies the same rules and thresholds to both backbones.GPT-5.6 Luna uses default API settings; OpenResearcher-30B-A3B is served locally with vLLM and BF16.Source: §B.1 · p. 12, B Implementation Details · p. 12

B.2 Source Discovery

• Citation fields determine source type and retrieval route; retrieval expands from direct and structured sources to broader web search when needed.For the GPT-5.6 Luna run, 2,913 of 3,356 citations (87%) required no commercial search, while query revision was triggered for 350.Source: §B.2 · pp. 12–13

B.3 Evidence Verification

• Field-level verification combines matches, conflicts, and missing evidence, while LLM judgments address ambiguous cases and are resolved by majority vote across three samples.Author mismatches require more reliable evidence; venue, year, or identifier discrepancies alone are not treated as contradictions.Source: §B.3 · pp. 13–14

B.4 Source Resolution

• Source-resolution candidates are grouped by shared work identifiers or compatible metadata, and require title overlap, identity evidence, and sufficient bibliographic information.Conflicting fields can be corrected rather than disqualifying a candidate when other evidence identifies the intended work.Source: §B.4 · pp. 14–15

B.3 Evidence Verification

• 2,692 of 3,356 citations (80%) received final decisions from deterministic rules; the LLM judge resolved the remaining 664 (20%).Table 9 summarizes these decision patterns.Source: §B.3 · p. 14, Table 9 · p. 15

B.4 Source Resolution

• After at most one targeted follow-up search, source resolution returns a corrected citation only when exactly one work is supported; otherwise it returns None.The returned citation fields come from retrieved records.Source: §B.4 · p. 15

B.5 Baseline Implementations

• The baselines include citation-only Direct Prompting, a same-backbone CiteAudit rerun, and Web Search, which adds search access to the source-resolution prompt.The CiteAudit rerun retains released prompts while replacing its extractor and judge with the compared backbone.Source: §B.5 · pp. 15–16

C Evaluation Details

• End-to-end success requires detecting a Fake citation and recovering its correct source or returning None when unrecoverable; predicted sources are checked field by field.Ambiguous fields are judged without revealing the original citation or the system that produced the prediction.Source: C Evaluation Details · p. 16

D Instruction Templates

• The appendix organizes instruction templates for CITEANCHOR’s three modules, ablations and baselines, and the evaluation judge, specifying each task’s inputs and output format.Templates also state the corresponding decision criteria.Source: D Instruction Templates · p. 16

D.1 CITEANCHOR

• CITEANCHOR’s templates separate citation-field extraction, source-type classification, and evidence verification, with instructions defining source categories and error criteria.The verification prompt uses the citation, retrieved record, and field comparisons; it distinguishes errors from acceptable version or metadata differences.Source: §D.1 · pp. 16–17
• The evidence-verification examples contrast a subtitle variation judged consistent with an author conflict flagged as an error despite an arXiv record identifying the intended work.The first example’s citation adds a subtitle; the second has the same title and identifier but conflicting authors.Source: §D.1 · p. 17

D.2 Ablation and Baselines Instruction for Single-Call Judgment

• Baseline templates distinguish citation-only judgments from record-assisted resolution, with both defining correct, recoverable, and unrecoverable outcomes.Web Search may search first, while Direct Prompting answers from its own knowledge.Source: §D.2 · p. 17

D.3 Evaluation

• The evaluation judge assesses predicted metadata fields against gold records without seeing the original citation or producing system, allowing version-aware title, author, venue, and year judgments.The judge accepts title variations such as added or dropped subtitles and author-list truncation with “et al.”Source: §D.3 · p. 18

E.1 Published and Rerun CiteAudit Results

• CITEANCHOR with GPT-5.6 Luna reached 0.841 F1, similar to the published CiteAudit result of 0.838, but flagged fewer valid citations and missed more hallucinated citations.The published result was 0.838 F1; CITEANCHOR flagged 54 valid citations and missed 89 hallucinated citations, versus 149 and 23 for that reference.Source: §E.1 · p. 18

E.2 Resolution Outcomes by Recoverability

• 74.7% task success for CITEANCHOR versus 58.4% for Web Search with GPT-5.6 Luna combines accurate recoverable-source proposals with abstention on unrecoverable citations.CITEANCHOR withholds correction for 92.4% of Unrecoverable cases, compared with 31.0% for Web Search.Source: §E.2 · p. 18

E.3 The Open-Weight Web Search Baseline

• 70.7% task success for CITEANCHOR with OpenResearcher-30B-A3B contrasts with Web Search’s 5.4%; CITEANCHOR takes corrected source details from retrieved records.Web Search often retained erroneous cited authors or proposed sources for Unrecoverable citations.Source: §E.3 · p. 18