CiteANCHORarXiv:2610.30355 · 2026

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

Chenyi Tong1, Ping Nie2, Yu Zhang3
1University of Wisconsin–Madison, 2University of Waterloo, 3Texas A&M University
Paper Code 🤗CiteResolve Benchmark
Figure 1: overview of CiteANCHOR with a running example

Open Figure 1 at full size

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.

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.

The CiteResolve Benchmark

Automated reference checks are already used at major conferences, with flagged references passed to human reviewers. For such screening to be useful, a system must explain why a citation is flagged and provide source records that allow reviewers to check its decision. Two challenges arise.

Interpreting heterogeneous source evidence

Papers, books, and web resources differ in how they can be retrieved and which fields identify them. A valid citation may omit a field present in a database record, while an invalid citation may match a record that omits the erroneous field. Detection needs source-aware interpretation that separates genuine contradictions from missing information and legitimate publication variants.

Separating citation errors from source identity

A citation with incorrect authors may still refer to an identifiable paper, whereas one with a fabricated title may have no real source at all. Evidence that establishes an error does not necessarily establish a replacement. A reliable system must decide whether a citation is wrong and, separately, whether its work can be uniquely identified, and abstain when it cannot.

CiteResolve makes this distinction measurable. Building on the Fake subset of the CiteAudit real-world test set, we annotate whether each hallucinated citation can be traced to a unique real work (Recoverable, with a verified corrected citation) or not (Unrecoverable, target None). Systems receive only the citation string and must decide whether it is hallucinated and, if so, return a uniquely supported source or None. Predicting Real on this subset always counts as a failure, so missed detections are penalized in the end-to-end score.

Cited asRecoverableUnrecoverableTotal
Paper225163388
Book314
Web resource†24731
Total252171423
What a recoverable citation gets wrongCitations
Authors only185 (73%)
Title only8 (3%)
Title and authors33 (13%)
Publication details15 (6%)
Web page details11 (4%)

Table 1: CiteResolve at a glance. †Web pages, datasets, software, and technical reports. Publication details refer to venue, year, volume, pages, or link. Web page details involve the page title, the author or organization, and the address.

Figure 2: additional statistics of CiteResolve

Figure 2: 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.

Source-Aware Grounding with CiteANCHOR

CiteANCHOR interprets retrieved records using source-specific criteria for citation verification and source resolution. Its central mechanism is evidence decomposition: individual bibliographic fields are checked before separate decisions are made about citation validity and source identity, so that a matching field does not validate the entire citation and an erroneous field does not automatically rule out its source. A chain-of-source strategy connects three modules that share the retrieved records:

  • Source discovery extracts title, authors, identifiers, venue, and year, determines whether the citation refers to a paper, book, or web resource, and retrieves records from cached and local resources first, then identifiers and structured bibliographic services (Crossref, OpenAlex, arXiv, Semantic Scholar, Europe PMC, Google Books, DataCite), and only then from Google Search and Google Scholar.
  • Evidence verification compares each field against the retrieved records with deterministic checks that distinguish contradictions from missing information and legitimate publication variants; an LLM judge handles only the comparisons these checks do not settle.
  • Source resolution reuses the collected evidence to recover the underlying work when exactly one record is supported, adds at most one targeted follow-up search when a detected error still has no supported source, and otherwise returns None.

The same retrieval rules and verification thresholds apply to both backbones we evaluate: the proprietary GPT-5.6 Luna and the open-weight OpenResearcher-30B-A3B, which supports local inference for confidential submissions.

Citation Detection on CiteAudit

All methods are compared on the same 3,356 real-world citations (467 hallucinated) of the CiteAudit test set. Published results provide context; the lower blocks compare methods using the same backbone, where CiteAudit is rerun with its released pipeline and prompts.

Confusion matrixMetrics
MethodTPFNFPTNAcc.Prec.Rec.F1
Published results for reference (Shi et al., 2026)
GPTZero3381291,3581,5310.5570.1990.7240.313
Gemini-3-Pro3511164122,4770.8430.4600.7520.571
GPT-5.23661011,3791,5100.5590.2100.7840.331
Claude-Sonnet-4.53491187562,1330.7400.3160.7470.444
CiteAudit (Gemini-3-Flash)444231492,7400.9490.7490.9510.838
Backbone: GPT-5.6 Luna
Direct Prompting120347542,8350.8810.6900.2570.374
CiteAudit3611061202,7690.9330.7510.7730.762
CiteANCHOR37889542,8350.9570.8750.8090.841
Backbone: OpenResearcher-30B-A3B
Direct Prompting1003674972,3920.7430.1680.2140.188
CiteAudit2222453792,5100.8140.3690.4750.416
CiteANCHOR384831282,7610.9370.7500.8220.784

Table 2: Detection on the CiteAudit real-world test set, with Fake as the positive class. Published results are taken from Shi et al. (2026) and not rerun; CiteAudit's published result uses Gemini-3-Flash. Bold marks the best metric within each backbone block.

Source-aware grounding improves both precision and recall over CiteAudit with the same backbone. With GPT-5.6 Luna, F1 rises from 0.762 to 0.841 (+7.9), with precision 0.875 vs. 0.751 and recall 0.809 vs. 0.773.
The gain is largest with the open-weight backbone. With OpenResearcher-30B-A3B, F1 rises from 0.416 to 0.784: false positives fall from 379 to 128 and missed hallucinated citations from 245 to 83. Deterministic field checks leave the backbone a narrow role, so the result depends less on which model fills it.

Source Resolution on CiteResolve

Under the end-to-end protocol, a system succeeds on a citation only if it flags it as hallucinated and returns the correct source (Recoverable) or None (Unrecoverable). Identified is the fraction of recoverable citations flagged with a proposed source, or of unrecoverable citations flagged with None; Recovered is the fraction of proposed sources that match the gold record on title, authors, venue, and year; Task success covers all 423 citations.

Recoverable (252)Unrecoverable (171)Overall
MethodIdentified ↑Recovered ↑Identified ↑Task success ↑
Backbone: GPT-5.6 Luna
Direct Prompting23.036.219.312.8
Web Search85.390.231.058.4
CiteANCHOR65.995.292.474.7
Backbone: OpenResearcher-30B-A3B
Direct Prompting25.812.33.53.3
Web Search61.910.93.55.4
CiteANCHOR68.787.986.070.7

Table 3: Source resolution on CiteResolve under the end-to-end protocol. All values are percentages. Web Search uses the same prompt as Direct Prompting but may search the web (OpenAI's web-search tool for GPT-5.6 Luna; OpenResearcher's browser tools for the open-weight backbone). Bold marks the best value within each backbone block.

Web Search proposes more sources; CiteANCHOR proposes the right ones and knows when to stop. With GPT-5.6 Luna, Web Search names a source for 85.3% of recoverable citations against 65.9% for CiteANCHOR, but CiteANCHOR's proposals are more accurate (95.2% vs. 90.2%) and it correctly withholds a correction on 92.4% of unrecoverable citations against 31.0%. The result is 74.7% task success, 16.3 points above Web Search.
Web access alone does not make an open-weight backbone a reliable resolver. With OpenResearcher-30B-A3B, CiteANCHOR is best on all four metrics and reaches 70.7% task success, against 5.4% for Web Search and 3.3% for Direct Prompting: the baseline tends to copy the cited title and author list back as the answer, while CiteANCHOR reads authors and source from the retrieved records.
Figure 3: commercial search use and estimated API cost

Figure 3: Commercial search use and estimated API cost over the 3,356 CiteAudit citations. (a) Number of SerpApi searches, including additional searches during source resolution: 3,204 → 1,039 with GPT-5.6 Luna and 4,071 → 1,066 with OpenResearcher-30B-A3B. (b) Estimated combined cost of LLM calls and SerpApi searches at public API prices; hatching marks the search share. With GPT-5.6 Luna the estimated cost per run falls from about $36 to $12.

Ablation Study

With GPT-5.6 Luna, we examine three design choices: evidence decomposition, complementary retrieval sources, and source-type adaptation. Each variant is evaluated on detection (CiteAudit F1) and on end-to-end resolution (CiteResolve task success).

VariantCiteAudit F1CiteResolve success (%)
w/o Chain (one LLM judgment over the same records)0.64359.3
w/o Search Engines (no Google and Google Scholar)0.73472.6
w/o Structured APIs (search engines and cited webpages only)0.77353.2
w/o Type Adaptation (generic rules for all source types)0.76166.9
CiteANCHOR0.84174.7

Table 4: Ablations with GPT-5.6 Luna. w/o Chain replaces staged verification and resolution with one LLM judgment over the same retrieved 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.

Evidence decomposition matters even with the retrieved evidence held fixed. Replacing the staged procedure with a single judgment over the same records lowers detection F1 by 19.8 points and resolution success by 15.4 points: a single judgment answers more often and is wrong more often, especially on unrecoverable citations.
Search engines and structured bibliographic records play complementary roles. Removing search engines mainly hurts detection (F1 0.841 → 0.734), while removing structured APIs mainly hurts resolution (success 74.7 → 53.2).
Source-type adaptation avoids losses on every kind of work. Generic rules reduce F1 by 8.0 points and success by 7.8 points; they flag far more real books and web resources, while paper-specific rules applied to everything would flag real web resources instead.

Case Studies

Verification must consider both what a retrieved record says and whether it describes the cited work itself.

Figure 4: two CiteResolve cases where all baselines fail

Figure 4: 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: seven of the ten cited authors are absent from the paper's complete 13-author list, yet all three baselines accept the citation. CiteANCHOR detects the contradiction and recovers the source. (b) An unrecoverable citation is repeated on the web without a supporting bibliographic record, and its arXiv identifier resolves to a different paper. CiteANCHOR rejects the mismatched identifier and returns None.

Code and Data

CiteANCHOR code

The framework, the run script, and the scorers for detection and source resolution.

CiteResolve benchmark

423 hallucinated citations with recoverability labels and verified corrected citations, on Hugging Face.

Open-weight backbone

CiteANCHOR also runs on a locally served open-weight model, so model inference stays local.

ResourceContentLicense
CiteANCHORsource discovery, evidence verification and source resolution modules; scripts that run the pipeline and compute the paper's numbersMIT
cheney777/CiteResolvethe 423 hallucinated citations with recoverability labels, gold sources and corrected citationsCC BY 4.0

BibTeX

arXiv:2610.30355 · cs.CL · 2026

@misc{tong2026citeanchorsourceawaregroundinghallucinated,
      title={CiteAnchor: Source-Aware Grounding for Hallucinated Citation Detection and Resolution},
      author={Chenyi Tong and Ping Nie and Yu Zhang},
      year={2026},
      eprint={2610.30355},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.30355},
}

University of Wisconsin–Madison University of Waterloo Texas A&M University