CiteAnchor 提出来源感知框架，检测并还原虚构引用

Ping Nie Affiliation: University of Waterloo Email: yuzhang@tamu.edu Yu Zhang Affiliation: 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.

1 Introduction

Large language models (LLMs) are increasingly integrated into the practice of scientific paper writing (Liang et al., 2025; Kusumegi et al., 2025). When these models generate references, however, they can invent papers or mix correct and incorrect details about real ones. Such hallucinated citations have already been found in published conference papers and across the scientific literature (Sakai et al., 2026a; Xu et al., 2026; Zhao et al., 2026). Their harms are evident: they can lend credibility to unsupported claims, spread misinformation through the scientific literature, contaminate the scientific knowledge base, and undermine decisions in high-stakes domains.

Interpreting Heterogeneous Source Evidence. Papers, books, and web resources differ in how they can be retrieved and which bibliographic fields identify them. Even for the same work, databases may omit different fields, while preprint and published versions may differ in year or venue (Visser et al., 2021; Shi et al., 2026). These differences make simple matching unreliable: 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. Reliable detection therefore requires source-aware interpretation of the evidence, distinguishing 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 an entirely fabricated title may have no real source at all. Both should be classified as hallucinated, but evidence that establishes an error does not necessarily establish a replacement. Recovering the correct source makes the discrepancy explicit and gives reviewers evidence to inspect. Selecting a merely plausible source, however, can introduce another false claim. A reliable system must therefore distinguish whether a citation is wrong from whether its underlying work can be uniquely identified, and abstain from correction when that identity remains unresolved.

To evaluate this distinction, we introduce the CiteResolve benchmark for source resolution in real-world hallucinated citations. Building on CiteAudit (Shi et al., 2026), we annotate recoverability and provide verified corrected references for citations with a uniquely supported source. CiteResolve evaluates correct recovery and appropriate abstention in an end-to-end setting, where systems receive only the original citation and must determine whether it is hallucinated and, if so, identify a uniquely supported source or conclude that no unique source can be established.

We further propose CiteAnchor, an agentic framework built on source-aware grounding: interpreting retrieved records using source-specific criteria for citation verification and source resolution. Its central mechanism is evidence decomposition: checking individual bibliographic fields before making separate decisions about citation validity and source identity. A chain-of-source strategy implements this decomposition through type-specific retrieval, deterministic bibliographic checks, and LLM reasoning for cases that these checks do not settle. Retrieved evidence is shared across discovery, verification, and resolution, with a targeted follow-up search when a detected error still has no supported source. This design grounds both hallucination labels and proposed corrections in verifiable records.

We evaluate CiteAnchor with the proprietary GPT-5.6 Luna (OpenAI, 2026a) and the open-weight OpenResearcher-30B-A3B (Li et al., 2026b). On real-world hallucinated citation detection, CiteAnchor improves F1 over the CiteAudit baseline (Shi et al., 2026) with the same backbones by 7.9 and 36.8 points, respectively. Even with the open-weight backbone, which supports local inference for confidential submissions, it exceeds the direct-prompting results reported by Shi et al. (2026) for Gemini-3-Pro (Google DeepMind, 2026), GPT-5.2 (OpenAI, 2025), and Claude-Sonnet-4.5 (Anthropic, 2025). On our CiteResolve benchmark, overall resolution success reaches 74.7% and 70.7%, exceeding the corresponding Web Search baselines by 16.3 and 65.3 percentage points. With the proprietary backbone, replacing our staged procedure with a single LLM judgment over the same retrieved evidence reduces detection F1 by 19.8 points, while removing source-type adaptation reduces it by 8.0 points. These ablations support the importance of evidence handling beyond retrieval alone. The full workflow also uses 70.7% fewer commercial search API calls on average across the two backbones than CiteAudit, corresponding to an estimated saving of approximately $25 in search credits per run over 3,356 citations.

Our contributions are threefold:

•
Conceptually, we extend hallucinated citation evaluation to source identity with the CiteResolve benchmark, measuring correct recovery and justified abstention against verified targets.
•
Methodologically, we introduce source-aware grounding through evidence decomposition, combining type-specific retrieval, explicit field checks, and separate validity and resolution decisions in our CiteAnchor framework.
•
Empirically, we show stronger detection and overall source resolution across proprietary and open-weight backbones, supported by controlled ablations and substantially fewer commercial search API calls.

2 Related Work

Hallucinated References and Bibliographic Verification. Hallucinated references occur in both LLM-generated texts (Walters and Wilder, 2023) and published papers (Sakai et al., 2026a; Xu et al., 2026; Zhao et al., 2026), while uneven database coverage and incomplete records complicate verification (Visser et al., 2021). CiteAudit (Shi et al., 2026) addresses this problem through multi-stage bibliographic lookup, web search, and evidence-based judgments. CiteTracer (Li et al., 2026a) combines a retrieval cascade with deterministic field checks and LLM judgments, distinguishing missing information from contradictions. CiteCheck (Khajavi et al., 2026) also identifies candidate publications before checking their fields and distinguishes references to identifiable works with corrupted metadata from those without credible matches, evaluating these labels on synthetic corruptions. HALLMARK (Reizinger and Brendel, 2026) evaluates verification under metadata errors and highlights false positives as a deployment concern. Our work builds on these retrieval and verification strategies through source-aware evidence decomposition, and extends evaluation to the correctness of recovered sources and abstention decisions against verified targets.

Citation Resolution and Correction. In bibliometrics, reference matching consolidates noisy variants of the same work within a dataset through normalization and metadata reconciliation (Aria et al., 2026). Scientific citation attribution starts from the surrounding text: CiteME (Press et al., 2024) evaluates recovery of author-cited papers, while CiteGuard (Choi et al., 2026) also identifies alternative valid citations. In retrieval-augmented generation, ALCE (Gao et al., 2023) evaluates citation support in generated answers, and CiteFix (Maheshwari et al., 2025) corrects citation assignments against retrieved documents. Our setting starts from the bibliographic entry itself, without the surrounding claim: the system must check whether it is valid and determine whether its remaining information identifies a unique real source.

3 The CiteResolve Benchmark

Detecting an incorrect citation and establishing its source require different evidence. CiteResolve makes this distinction measurable by pairing real-world hallucinated citations with verified source-resolution targets. We evaluate detection and resolution together, so that success requires identifying the error and correctly determining what source, if any, can be recovered.

3.1 Benchmark Construction

We build on the real-world test set of CiteAudit (Shi et al., 2026), denoted by 𝒞={(ci,li)}i=1N, where ci is a citation and li∈{Real,Fake} is its label. Real denotes a citation that is valid as written, whereas Fake covers nonexistent works and substantive errors in comparison with real works. We retain the full dataset and its released labels for evaluating hallucinated citation detection.

To evaluate source resolution, we further annotate CiteAudit’s Fake subset. We first determine whether a citation can be traced to a unique real work. If so, we label it Recoverable and provide a verified corrected citation ci⋆. Otherwise, we label it Unrecoverable and set ci⋆=None. (Note that CiteAudit provides neither these recoverability labels nor the corrected citations.) Let ℐ⊆{i:li=Fake} index the cases retained after manual review and input filtering. The resulting benchmark, 𝒞~={(ci,ci⋆)}i∈ℐ, contains 423 citations: 252 Recoverable and 171 Unrecoverable. Table 1 summarizes their source types and error patterns.

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.

3.2 Annotation Criteria

Annotation separates evidence of an error from evidence of source identity. Following reference-matching studies (Visser et al., 2021; Han et al., 2026), we examine title similarity together with author and identifier evidence, allowing minor formatting differences. The criteria vary with the type of source being cited.

Papers. A near-exact title match is the main basis for recovery, allowing differences in capitalization, punctuation, and minor typographical errors. When the title is very close in meaning but not a near-exact match, we additionally require an identifier in the citation, such as an arXiv ID or URL, to point to the candidate work. An incorrect author list does not prevent recovery if the remaining evidence uniquely identifies that work. Year differences alone do not determine recoverability, since online and print publication dates can differ (Aria et al., 2026). Note that these rules concern whether the source can be recovered. Incorrect fields still need to be corrected in c⋆.

Books and Web Resources. For books, we require agreement in both title and authors. Minor publisher discrepancies do not by themselves rule out recovery. For web resources, we follow the cited URL and compare the page title and any available author or organization information. A page without an identifiable author is not rejected for that reason alone, consistent with reference conventions for works without author information³³ 3 https://apastyle.apa.org/style-grammar-guidelines/references/missing-information. An explicit author mismatch, however, remains a discrepancy to resolve.

Unrecoverable Citations. We assign c⋆=None when no candidate satisfies these criteria or when several distinct works remain plausible. Thus, None denotes the absence of a uniquely supported correction under our criteria, rather than a claim that no related work exists.

3.3 Evaluation Protocol

We run the evaluated algorithm on every citation ci in 𝒞, providing only the citation string. It first predicts a label l^i and, when l^i=Fake, determines whether the citation can be recovered. If it judges the citation Recoverable, it returns a proposed corrected citation c^i∗. If it judges the citation Unrecoverable, it returns c^i∗=None.

Detection is evaluated over all of 𝒞 against li, with accuracy defined as

1|𝒞|∑i=1|𝒞|𝟏[l^i=li].

(1)

Resolution is evaluated over the citations indexed by ℐ. A successful outcome requires both a correct Fake prediction and a correct recovery or abstention:

1|ℐ|∑i∈ℐ𝟏[l^i=Fake]Match(c^i∗,ci⋆).

(2)

Here, Match⁡(c^i∗,ci⋆)=1 if the predicted correction is judged correct under the procedure in Section 5.1, or if both values are None. Otherwise, it is 0. Predicting Real always counts as a resolution failure on this subset, so missed detections are penalized in the end-to-end score.

4 The CiteAnchor Framework

CiteAnchor implements source-aware grounding by tailoring retrieval tools and bibliographic checks to each source type. Its evidence decomposition separates field comparisons, 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. Inspired by the chain-of-evidence principle (Meng et al., 2026), our chain-of-source strategy connects three modules: source discovery 𝖣θ, evidence verification 𝖵θ, and source resolution 𝖱θ. They combine deterministic checks with LLM reasoning and share retrieved records, where θ denotes the common backbone used for their LLM calls. Figure 1 gives an overview of CiteAnchor.

Refer to caption — 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.

4.1 Source Discovery

Given a citation c, the source discovery module 𝖣θ extracts its title, authors, identifiers, venue, and year when available. It uses identifiers for direct lookups and title/author information to construct search queries. It also identifies whether the citation refers to a paper, book, or web resource, which determines the relevant retrieval tools. Starting with ℰ0=∅, 𝖣θ retrieves records ℛt and accumulates evidence ℰt:

ℛt=𝖣θ(c,ℰt−1),ℰt=ℰt−1∪ℛt.

(3)

Retrieval first checks cached records and local resources, such as an offline copy of the ACL Anthology (Bird et al., 2008). It then follows available identifiers and queries structured bibliographic services, such as Crossref (Hendricks et al., 2020). Broader web search, such as Google Search (through SerpApi⁴⁴ 4 https://serpapi.com/search-api), is used only when these sources fail to provide a plausible matching record. The concrete source configuration is given in Section 5.1.

When incorrect fields obscure a match, 𝖣θ can revise the query, for example by searching by title after author information leads to conflicting results.

4.2 Evidence Verification

Given c and the retrieved evidence ℰ, the evidence verification module 𝖵θ compares the cited fields with available records, distinguishing agreement, contradiction, and missing evidence. If a retrieved record omits a field, it cannot confirm whether the corresponding value in the citation is correct or not. Conversely, omitting an optional field from the citation is not itself an error. The comparison also accounts for source type and legitimate year or venue differences between publication versions (Visser et al., 2021; Shi et al., 2026). Deterministic checks handle clear matches and contradictions, and the LLM assesses cases that these checks do not settle. The resulting validity decision is

l^=𝖵θ(c,ℰ),l^∈{Real,Fake}.

(4)

For citations predicted as Fake, the evidence is passed to source resolution to either recover a uniquely supported source or conclude that no unique source can be established.

4.3 Source Resolution

For a citation c predicted to be hallucinated, the source resolution module 𝖱θ uses the collected evidence ℰ to establish which real work, if any, remains supported. It groups records of the same work and represents each group by one candidate citation. The field comparisons distinguish evidence supporting a candidate’s identity from discrepancies that require correction. For example, a contradicted author list can be corrected when the title and other available clues still identify the work. Let 𝒰⁡(c,ℰ) denote the candidates that the evidence supports as possible sources of c. If this set is empty, the module makes one additional targeted web search through SerpApi. The returned records ℛextra are added to the evidence set: ℰ←ℰ∪ℛextra. The module then checks the candidates against all available evidence and returns c^∗=𝖱θ(c,ℰ), with

c^∗={x,𝒰⁡(c,ℰ)={x},None,otherwise.

(5)

A unique supported candidate becomes the proposed corrected citation. If no candidate is supported or several distinct works remain plausible, None withholds a correction while retaining the Fake judgment.

5 Experiments

5.1 Experimental Setup

Datasets. We evaluate detection on CiteAudit and source resolution on CiteResolve, following the end-to-end protocol in Section 3.3.

Backbones. To demonstrate that CiteAnchor supports different backbones, we instantiate it with the proprietary GPT-5.6 Luna (OpenAI, 2026a) and the open-weight OpenResearcher-30B-A3B (Li et al., 2026b). The latter is post-trained on multi-step search and browsing trajectories, making it a suitable choice for agentic information seeking. Including such an open-weight backbone also allows local model inference when submission confidentiality restricts the use of external LLM services.

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.

Baselines. For detection, we include the results reported by Shi et al. (2026) for GPTZero (Adam et al., 2026), Gemini-3-Pro (Google DeepMind, 2026), GPT-5.2 (OpenAI, 2025), Claude-Sonnet-4.5 (Anthropic, 2025), and CiteAudit (Shi et al., 2026) with the Gemini-3-Flash backbone (Google DeepMind, 2025). These published results serve as reference points and are not rerun. With both backbones, we run Direct Prompting, which answers without retrieval, and CiteAudit (Shi et al., 2026), which uses the released pipeline. For the CiteAudit reruns, we replace its original backbone while retaining the released prompts. For resolution, Direct Prompting predicts citation validity and returns a source or None for citations it flags. Web Search uses the same prompt but permits web search through OpenAI’s web-search tool for GPT-5.6 Luna (OpenAI, 2026a) or OpenResearcher’s browser tools (Li et al., 2026b). Implementations are provided in Appendix B.5.

Configuration. Retrieval first checks records cached within the current run and the offline ACL Anthology (Bird et al., 2008), then follows identifiers and queries source-specific services. Identifiers include DOIs, arXiv IDs, and cited URLs. The services include Crossref (Hendricks et al., 2020), OpenAlex (Priem et al., 2022), arXiv (Ginsparg, 2011), Semantic Scholar (Lo et al., 2020), Europe PMC (The Europe PMC Consortium, 2015), Google Books, and DataCite (DataCite Metadata Working Group, 2026). Google Scholar and Google Search through SerpApi are used when earlier sources provide no plausible match. Deterministic checks handle clear cases, while unresolved comparisons go to the LLM judge. Resolution reuses the collected evidence and adds at most one targeted search when no candidate is supported. The same retrieval rules and verification thresholds apply to both backbones. Appendices B.2–B.4 specify retrieval routes, decision rules, and inference settings.

Evaluation Metrics. For detection, following Shi et al. (2026), we report accuracy (as defined in Section 3.3), precision, recall, and F1, treating Fake as positive. For resolution, we report Identified (Recoverable): the fraction of the 252 recoverable citations predicted Fake with a proposed source. Recovered measures conditional recovery accuracy: among these proposals, the fraction that matches the gold source. Identified (Unrecoverable) is the fraction of the 171 unrecoverable citations predicted Fake with None. Task success (as defined in Section 3.3) measures correct detection followed by correct recovery or abstention across all 423 citations. Recovered sources are compared with gold records by title, authors, venue, and year, allowing formatting differences and legitimate publication variants. Deterministic rules resolve clear matches, while GPT-6 Sol (OpenAI, 2026b) judges ambiguous fields without access to the original citation or system identity. Appendix C details the matching procedure.

5.2 Citation Detection on the CiteAudit Real-World Test Set

Overall Results. Table 2 compares all methods on the same 3,356 citations. The published results provide context, while the lower blocks compare methods using the same backbone. With GPT-5.6 Luna, CiteAnchor improves F1 over CiteAudit from 0.762 to 0.841, a gain of 7.9 points. Both precision (0.875 vs. 0.751) and recall (0.809 vs. 0.773) improve. The gain is larger with OpenResearcher-30B-A3B, where F1 rises from 0.416 to 0.784. Here, CiteAnchor reduces false positives from 379 to 128 and missed hallucinated citations from 245 to 83. These improvements show that source-aware grounding benefits both proprietary and open-weight backbones.

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.

Efficiency and Cost. Source-aware retrieval also reduces reliance on commercial search. Over the 3,356 citations, SerpApi calls decrease from 3,204 to 1,039 with GPT-5.6 Luna and from 4,071 to 1,066 with OpenResearcher-30B-A3B, reductions of 67.6% and 73.8%, respectively (Figure 2). With GPT-5.6 Luna, the estimated combined cost of LLM calls and commercial searches falls from approximately $36 to $12 per run.

5.3 Source Resolution on CiteResolve

Overall Results. Table 3 evaluates both recovering a uniquely supported source and withholding a correction when no such source can be established. With GPT-5.6 Luna, Web Search proposes sources for more recoverable citations than CiteAnchor (85.3% vs. 65.9%). However, CiteAnchor’s proposed sources are more accurate (95.2% vs. 90.2%), and it correctly identifies far more unrecoverable citations (92.4% vs. 31.0%). This combination yields 74.7% overall task success, exceeding Web Search by 16.3 percentage points despite lower recovery coverage.

The advantage is larger with OpenResearcher-30B-A3B. CiteAnchor performs best on all four metrics and reaches 70.7% task success, compared with 5.4% for Web Search and 3.3% for Direct Prompting. These results show that access to web search alone does not ensure reliable source resolution with this backbone.

6 Analysis

6.1 Ablation Studies

We examine three design choices in CiteAnchor: evidence decomposition, complementary retrieval sources, and source-type adaptation. Table 4 reports detection F1 on CiteAudit and resolution success on CiteResolve with GPT-5.6 Luna.

Evidence Decomposition. We first compare the full verification and resolution procedure with a single joint judgment over the same retrieved records. The w/o Chain variant asks the backbone to decide citation validity and source identity in one call, using the prompt in Appendix D.2. Detection F1 falls from 0.841 to 0.643, and resolution success from 74.7% to 59.3%. The comparison evaluates the decision procedure as a whole, including its field checks and instructions. It supports the benefit of explicitly checking citation fields and separating validity from source identity, even when the retrieved evidence is held fixed.

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.

Retrieval Sources. Removing search engines lowers detection F1 from 0.841 to 0.734, while resolution success decreases more modestly, from 74.7% to 72.6%. Removing structured APIs also hurts both tasks, but the larger loss is in resolution: success falls to 53.2%, while detection F1 remains at 0.773. The different patterns suggest complementary roles for the two source families. Broad web search helps detect citation errors, while structured bibliographic records are particularly valuable for establishing which source can be recovered.

Source-Type Adaptation. Finally, replacing type-specific retrieval and verification with a generic procedure reduces detection F1 to 0.761 and resolution success to 66.9%. Papers, books, and web resources differ in where they can be found and what evidence identifies them. These decreases support adapting both evidence collection and interpretation to the source type.

6.2 Case Studies

Figure 3 illustrates why verification must consider both what a retrieved record says and whether it describes the cited work itself. In Case (a), the title identifies a real paper, but seven of the ten cited authors are absent from its complete 13-author list. All three baselines accept the citation. CiteAudit attributes the discrepancy to truncated or differing records, while Web Search and Direct Prompting incorrectly state that the author list matches. CiteAnchor detects the author contradiction while using the matching title to recover the source. The same evidence thus supports a Fake label and a recoverable source.

In Case (b), the citation appears on the web, but the retrieved bibliographic records do not establish a matching source. Its arXiv identifier resolves to a different paper. All three baselines again accept the citation, and CiteAudit’s explanation specifically cites a webpage reproducing its title and authors. CiteAnchor rejects the mismatched identifier and returns Fake with None. This case illustrates why a webpage that repeats a reference is insufficient evidence that the cited work exists.

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.

7 Conclusion

We introduce CiteResolve, a benchmark for source resolution in hallucinated citations, and CiteAnchor, a source-aware grounding framework that separates citation validity from source identity. Across proprietary and open-weight backbones, CiteAnchor improves hallucinated citation detection and end-to-end source resolution while reducing commercial search use. Our results suggest that reliable citation verification should go beyond detecting errors to recover the underlying source when uniquely supported, and abstain from correction otherwise. This helps human reviewers inspect detected errors and assess proposed corrections using evidence from retrieved records.

Limitations

First, CiteResolve is dominated by citations to papers, with fewer books and web resources. The reported results therefore provide stronger evidence for paper citations than for other source types. Broader coverage across disciplines and languages would also help assess how well the framework transfers to other settings. Second, our task concerns bibliographic validity and source identity. The system receives only a citation entry and does not assess whether the cited work supports the surrounding claim. A correct reference can still be used in an unsupported context, so bibliographic verification and claim-level attribution should be viewed as complementary tasks. Finally, our evaluation reflects the records and services available at the time of the experiments. Bibliographic databases and web pages change, and the same citation may become easier or harder to resolve as records are added or revised. Accordingly, Unrecoverable means that no unique source is supported by the available evidence under our criteria, rather than that no related work exists.

Ethical Considerations

CiteAnchor is intended to help reviewers detect and examine problematic references. It is not intended to let authors bypass reference checking or rely on automatic corrections without verification. Authors remain responsible for the accuracy of every reference they include. That said, a Fake prediction by CiteAnchor concerns the citation itself and does not establish LLM use, deliberate fabrication, or author misconduct (Sakai et al., 2026a).

Our experiments use citations from the existing CiteAudit benchmark and public bibliographic sources. Deployment on confidential submissions requires model and retrieval services that meet the venue’s confidentiality requirements. A locally hosted backbone (e.g., OpenResearcher-30B-A3B (Li et al., 2026b) used in our experiments) can keep model inference local, but external searches still transmit queries to service providers. Only the information needed for reference verification should be shared, subject to the applicable confidentiality rules and service terms. We used AI assistants for coding and for writing assistance.

References

Adam et al. (2026) G. A. Adam, A. Cui, E. Thomas, E. Napier, N. Shmatko, J. Schnell, J. J. Tian, A. Dronavalli, E. Tian, and D. Lee GPTZero: robust detection of LLM-generated texts. Note: arXiv preprint arXiv:2602.13042 External Links: 2602.13042, Document, Link Cited by: §5.1, Table 2.
Anthropic (2025) Anthropic Claude Sonnet 4.5 system card. Technical report Anthropic. External Links: Link Cited by: §1, §5.1, Table 2.
Aria et al. (2026) M. Aria, L. D’Aniello, and M. Spano A multi-phase reference matching algorithm for bibliometric analysis: design, implementation, and evaluation. Scientometrics 131 (9), pp. 6693–6728. External Links: Document, Link Cited by: §2, §3.2.
Bird et al. (2008) S. Bird, R. Dale, B. J. Dorr, B. Gibson, M. T. Joseph, M. Kan, D. Lee, B. Powley, D. R. Radev, and Y. F. Tan The ACL Anthology Reference Corpus: a reference dataset for bibliographic research in computational linguistics. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08), pp. 1755–1759. External Links: Link Cited by: §4.1, §5.1.
Choi et al. (2026) Y. M. Choi, X. Guo, Y. R. Fung, and Q. Wang CiteGuard: faithful citation attribution for LLMs via retrieval-augmented validation. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 6241–6257. External Links: Document, Link Cited by: §2.
DataCite Metadata Working Group (2026) DataCite Metadata Working Group DataCite metadata schema documentation for the publication and citation of research data and other research outputs. DataCite e.V.. Note: Version 4.7 External Links: Document, Link Cited by: §5.1.
Gao et al. (2023) T. Gao, H. Yen, J. Yu, and D. Chen Enabling large language models to generate text with citations. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 6465–6488. External Links: Document, Link Cited by: §2.
Ginsparg (2011) P. Ginsparg ArXiv at 20. Nature 476 (7359), pp. 145–147. External Links: Document, Link Cited by: §5.1.
Google DeepMind (2025) Google DeepMind Gemini 3 Flash model card. Technical report Google DeepMind. External Links: Link Cited by: §5.1, Table 2.
Google DeepMind (2026) Google DeepMind Gemini 3 Pro model card. Technical report Google DeepMind. Note: Model released November 2025; model card updated May 2026 External Links: Link Cited by: §1, §5.1, Table 2.
Han et al. (2026) C. Han, S. Manoharan, X. Ye, and U. Speidel Phantom citations: an empirical study of non-existent and unverifiable references in scholarly literature. Journal of the Association for Information Science and Technology 77 (10), pp. 1287–1316. External Links: Document, Link Cited by: §3.2.
Hendricks et al. (2020) G. Hendricks, D. Tkaczyk, J. Lin, and P. Feeney Crossref: the sustainable source of community-owned scholarly metadata. Quantitative Science Studies 1 (1), pp. 414–427. External Links: Document, Link Cited by: §4.1, §5.1.
Khajavi et al. (2026) K. Khajavi, S. Sadeghi, R. Adhikari, and A. Tessier CiteCheck: retrieval-grounded detection of LLM citation hallucinations in scientific text. Note: arXiv preprint arXiv:2605.27700 External Links: 2605.27700, Document, Link Cited by: §1, §2.
Kusumegi et al. (2025) K. Kusumegi, X. Yang, P. Ginsparg, M. de Vaan, T. Stuart, and Y. Yin Scientific production in the era of large language models. Science 390 (6779), pp. 1240–1243. External Links: Document, Link Cited by: §1.
Li et al. (2026a) M. Li, Z. Lin, and S. Ma Source or it didn’t happen: a multi-agent framework for citation hallucination detection. Note: arXiv preprint arXiv:2605.08583 External Links: 2605.08583, Document, Link Cited by: §1, §2.
Li et al. (2026b) Z. Li, D. Jiang, X. Ma, H. Zhang, P. Nie, Y. Zhang, K. Zou, J. Xie, Y. Zhang, and W. Chen OpenResearcher: a fully open pipeline for long-horizon deep research trajectory synthesis. Note: arXiv preprint arXiv:2603.20278 External Links: 2603.20278, Document, Link Cited by: §B.5, §1, §5.1, §5.1, Table 2, Table 3, Ethical Considerations.
Liang et al. (2025) W. Liang, Y. Zhang, Z. Wu, H. Lepp, W. Ji, X. Zhao, H. Cao, S. Liu, S. He, Z. Huang, D. Yang, C. Potts, C. D. Manning, and J. Zou Quantifying large language model usage in scientific papers. Nature Human Behaviour 9 (12), pp. 2599–2609. External Links: Document, Link Cited by: §1.
Lo et al. (2020) K. Lo, L. L. Wang, M. Neumann, R. Kinney, and D. S. Weld S2ORC: the Semantic Scholar open research corpus. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 4969–4983. External Links: Document, Link Cited by: §5.1.
Maheshwari et al. (2025) H. Maheshwari, S. Tenneti, and A. Nakkiran CiteFix: enhancing RAG accuracy through post-processing citation correction. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track), pp. 310–317. External Links: Document, Link Cited by: §2.
Meng et al. (2026) R. Meng, B. Dalvi Mishra, J. Chen, C. Li, P. Goyal, M. Parmar, Y. Song, Y. Song, R. Sinha, P. Ranganathan, B. Gokturk, J. Yoon, and T. Pfister ScientistOne: towards human-level autonomous research via chain-of-evidence. Note: arXiv preprint arXiv:2605.26340 External Links: 2605.26340, Document, Link Cited by: §4.
OpenAI (2025) OpenAI Update to GPT-5 system card: GPT-5.2. Technical report OpenAI. External Links: Link Cited by: §1, §5.1, Table 2.
OpenAI (2026a) OpenAI GPT-5.6 Luna model. Note: Model documentationAccessed: 2026-10-08 External Links: Link Cited by: §1, §5.1, §5.1, Table 2, Table 3.
OpenAI (2026b) OpenAI GPT-6 Sol model. Note: Model documentationAccessed: 2026-10-08 External Links: Link Cited by: Appendix C, §5.1.
Press et al. (2024) O. Press, A. Hochlehnert, A. Prabhu, V. Udandarao, O. Press, and M. Bethge CiteME: can language models accurately cite scientific claims?. In Advances in Neural Information Processing Systems, Vol. 37, pp. 7847–7877. External Links: Document, Link Cited by: §2.
Priem et al. (2022) J. Priem, H. Piwowar, and R. Orr OpenAlex: a fully-open index of scholarly works, authors, venues, institutions, and concepts. Note: arXiv preprint arXiv:2205.01833 External Links: 2205.01833, Document, Link Cited by: §5.1.
Reizinger and Brendel (2026) P. Reizinger and W. Brendel HALLMARK: diagnosing three failure modes in LLM citation verifiers. Note: arXiv preprint arXiv:2607.18360 External Links: 2607.18360, Document, Link Cited by: §2.
Sakai et al. (2026a) Y. Sakai, H. Kamigaito, and T. Watanabe HalluCitation matters: revealing the impact of hallucinated references with 300 hallucinated papers in ACL conferences. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 47295–47376. External Links: Document, Link Cited by: §1, §2, Ethical Considerations.
Sakai et al. (2026b) Y. Sakai, H. Kamigaito, and T. Watanabe HalluCiteChecker: a lightweight toolkit for hallucinated citation detection and verification in the era of AI scientists. Note: arXiv preprint arXiv:2604.26835 External Links: 2604.26835, Document, Link Cited by: §1.
Shi et al. (2026) K. Shi, W. Sun, Z. Zhang, L. Sun, N. V. Chawla, and Y. Ye CiteAudit: you cited it, but did you read it? a benchmark for verifying scientific references in the LLM era. Note: arXiv preprint arXiv:2602.23452 External Links: 2602.23452, Document, Link Cited by: §1, §1, §1, §1, §2, §3.1, §4.2, §5.1, §5.1, Table 2, Table 2, Table 2, Table 2, Table 2.
The Europe PMC Consortium (2015) The Europe PMC Consortium Europe PMC: a full-text literature database for the life sciences and platform for innovation. Nucleic Acids Research 43 (D1), pp. D1042–D1048. External Links: Document, Link Cited by: §5.1.
Visser et al. (2021) M. Visser, N. J. van Eck, and L. Waltman Large-scale comparison of bibliographic data sources: Scopus, Web of Science, Dimensions, Crossref, and Microsoft Academic. Quantitative Science Studies 2 (1), pp. 20–41. External Links: Document, Link Cited by: §1, §2, §3.2, §4.2.
Walters and Wilder (2023) W. H. Walters and E. I. Wilder Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports 13 (1), pp. 14045. External Links: Document, Link Cited by: §2.
Xu et al. (2026) Z. Xu, Y. Qiu, L. Sun, F. Miao, F. Wu, X. Li, X. Wang, H. Lu, Z. Zhang, Y. Hu, J. Li, L. Jin, F. Zhang, R. Luo, X. Liu, Y. Li, and J. Liu GhostCite: a large-scale analysis of citation validity in the age of large language models. Note: arXiv preprint arXiv:2602.06718 External Links: 2602.06718, Document, Link Cited by: §1, §2.
Zhao et al. (2026) Z. Zhao, Y. Wang, T. Stuart, M. de Vaan, P. Ginsparg, and Y. Yin LLM hallucinations in the wild: large-scale evidence from non-existent citations. Note: arXiv preprint arXiv:2605.07723 External Links: 2605.07723, Document, Link Cited by: §1, §2.

Appendix Contents

Appendix A CiteResolve Details

A.1 Construction

Starting Point. CiteResolve is constructed from citations labeled Fake in the CiteAudit real-world test set. We retain the citation string from the original paper and manually annotate whether its source can be uniquely identified. After annotation and quality control, the benchmark contains 423 citations.

Each retained citation is labeled Recoverable when the available evidence uniquely identifies a source, or Unrecoverable when no unique source can be established.

Annotation Workflow. For each citation, the annotation tool first displayed the top three records returned by Crossref for the full reference string and by OpenAlex and Semantic Scholar for the parsed title. The returned records were pooled without reranking. If none matched, the annotator could broaden the query, resolve a DOI, arXiv ID, or URL, search by free text, or enter a source found elsewhere. CiteAnchor was not used during annotation. Annotation proceeded in two passes: one annotator labeled every citation that CiteAudit marks as Fake, and a second annotator independently re-annotated all of them under the same criteria; disagreements were resolved by discussion (see Quality Control below).

Source-Identity Criterion. The annotation decision is based on whether the evidence identifies a unique underlying source. For papers, title evidence carries the most weight; an author list alone is not sufficient to establish the intended work. Identifiers such as DOI, arXiv ID, and URL are first resolved and then checked against the remaining citation fields.

A citation is labeled Recoverable only when the available evidence points consistently to one source. If no compatible source can be found or multiple sources remain plausible, it is labeled Unrecoverable. For example, a partially corrupted title may match several real works while the author list points to another source. Without a unique match, the citation remains Unrecoverable.

Different versions of the same underlying work, such as an arXiv preprint and its later conference publication, are treated as the same source when their identity is clear. Minor bibliographic differences, including year differences across versions, a trailing “et al.” or “others” in the author list, or fields that are missing on either side, do not by themselves make a citation unrecoverable.

Gold Sources. For each Recoverable citation, we record the verified source’s title, authors, venue, year, and available identifier. This record serves as the gold target for evaluation.

For web resources without standard bibliographic metadata, the gold source is the resolved webpage, dataset page, or repository rather than a complete bibliographic record. We therefore leave unavailable fields empty and record the corrected source directly.

Quality Control. After the initial annotation, all citations were checked again for extraction and metadata errors. A second annotator independently re-annotated the citations using the same source-identity criteria, without access to the first-pass labels, and consulted the cited papers or authoritative web sources for uncertain cases. On the 423 citations in the benchmark, the second annotator assigned a different recoverability label to 8 citations and flagged 31 others as ambiguous under the criteria; the two passes agree on 415 labels (Cohen’s κ=0.96). After discussion, 6 labels were changed, the other 2 were kept, and the criteria were clarified for the ambiguous cases.

A.2 Data Analysis

Composition. Table 1 summarizes the 423 citations. Most citations refer to papers (388), followed by web resources (31) and books (4). The web resources include 17 webpages, 6 datasets, 6 reports, and 2 software repositories.

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.

Figure 4 provides a broader view of the benchmark. For the 252 Recoverable citations, most gold records come from OpenAlex, Crossref, or arXiv, with Google Scholar and cited webpages covering most of the remaining cases. The benchmark spans a range of research areas. Computing accounts for the largest share, while the remaining citations cover the natural sciences, medicine and health, and law, social sciences, and education. Most cited sources are journal or conference papers or arXiv and other unpublished works, with additional coverage of books, reports, websites, datasets, software, blogs, and news articles.

Identifiers. Table 5 shows that identifiers are common in both classes. However, a valid identifier does not necessarily identify the cited source. Among Unrecoverable citations, 35 of 39 arXiv IDs are valid, yet none of the resolved papers matches the cited title. Identifiers must therefore be resolved and checked against the remaining citation evidence.

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.

Appendix B Implementation Details

This appendix specifies how the modules of Section 4 are instantiated in the reported runs. Appendix B.1 gives the setup shared by all systems, Appendices B.2–B.4 follow source discovery, evidence verification, and source resolution, and Appendix B.5 describes the baselines. Statistics refer to the GPT-5.6 Luna run over the 3,356 CiteAudit citations.

B.1 Common Setup

Backbones and Serving. Within each backbone comparison block of Tables 2 and 3, all compared systems use the same backbone. GPT-5.6 Luna is called through the OpenAI API with its default sampling and reasoning settings and at most 4,096 output tokens per call, except for the CiteAudit rerun, which uses 8,192 (Appendix B.5). We serve OpenResearcher-30B-A3B locally with vLLM 0.13, as specified in the authors’ implementation, using BF16 on one NVIDIA H20 per server, with a 32k context window and prefix caching. For CiteAnchor, Direct Prompting, and the CiteAudit rerun, OpenResearcher’s native browser tools are disabled. Retrieval in CiteAnchor and CiteAudit is handled by their respective pipelines. OpenResearcher’s reasoning budget is capped at 2,048 tokens, and responses are limited to 4,096 tokens and constrained to the required JSON schema. CiteAnchor uses the same rules and thresholds for both backbones.

B.2 Source Discovery

Fields and source type. 𝖣θ extracts the title, authors, identifier, venue, and year of a citation with one backbone call. It then assigns a source type, which determines the retrieval route and the evidence used for verification. We first apply deterministic bibliographic cues in a fixed order (Table 6), covering papers, books, reports, web pages, datasets, and software repositories. The LLM backbone is used when these cues do not determine a type. Citations whose source type remains unresolved follow the paper route.

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.

Type-Specific Retrieval. The source type determines the retrieval route and the sources consulted (Table 7). Papers are searched through bibliographic services, books additionally through Google Books, and web resources first at their cited address. Within each route, retrieval proceeds from direct and structured sources to broader web search. The system first checks previously retrieved records and an offline copy of the ACL Anthology, then identifier lookups and structured bibliographic services, and uses Google Scholar or Google Search only when the earlier sources do not yield a plausible candidate. Retrieval broadens only when the current evidence does not provide a usable candidate. Overall, 2,913 of the 3,356 citations (87%) required no commercial search.

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.

Query Revision. An incorrect citation field can prevent the intended source from appearing in the initial search. When no plausible record is found, all retrieved candidates are ruled out as different works, or the remaining evidence is too weak to verify the citation, 𝖣θ revises the query by dropping or simplifying unreliable fields. For example, it may search by title alone or use a shortened title. The revised search stops once a plausible record is found. In the GPT-5.6 Luna run, query revision was triggered for 350 citations.

B.3 Evidence Verification

𝖵θ compares a citation with the retrieved records field by field, combines the resulting evidence into a validity decision, and applies additional checks to unreliable author metadata.

Field-Level Comparison. Each field receives one of three outcomes, corresponding to agreement, contradiction, and missing evidence in Section 4.2: match, conflict, and missing. missing means that the available evidence is insufficient to establish either agreement or contradiction. This can occur because a field is absent or because the comparison falls into an ambiguous range.

Titles are compared by word overlap after normalization. Let A and B be the token sets of the two titles. We use two measures:

overlap=|A∩B|max⁡(|A|,|B|),

containment=|A∩B|min⁡(|A|,|B|).

The first measures overall overlap, while the second captures cases where one title is largely contained in the other.

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 Rules. The field outcomes in Table 8 are combined to determine whether the citation is supported. Title and author evidence establishes the identity of the cited work. Differences in venue, year, or identifier are checked against legitimate versions of the work before being treated as errors.

Retrieved candidates are examined in decreasing order of title agreement, title similarity, year agreement, and author overlap. For each candidate, deterministic rules first handle clear cases. A citation is Real when the candidate supports the cited work without a conflict in its identity evidence. It is Fake when the evidence establishes an identity conflict, such as incorrect authors or an identifier resolving to a different work, or when all retrieval routes have been exhausted without finding a matching work. Cases not settled by these rules are passed to the LLM judge.

LLM Judgment for Ambiguous Cases. Cases not settled by the deterministic rules are passed to the LLM judge. These include borderline title or author matches, possible differences between publication versions, and conflicts supported by only one metadata source.

The LLM receives the parsed citation, one candidate record with its provenance, and the field-level comparison results, with the disputed fields highlighted (Appendix D). It returns error, consistent, or different work. Each judgment is sampled three times and resolved by majority vote. If a candidate is judged to refer to a different work, it is discarded and verification proceeds to the next candidate. The only exception is when the record matches the cited title and authors, in which case we treat it as another version of the same work. We act on an error verdict only when the title or authors conflict; discrepancies limited to the venue, year, or identifier are not treated as contradictions. This follows the same criterion used for the baselines (Appendix B.5).

Deterministic rules make the final decision for 2,692 of the 3,356 citations (80%), while the LLM judge resolves the remaining 664 (20%). Table 9 summarizes the resulting decision patterns.

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
Fake: no supported work is found
No supported record after retrieval	rules	173
All candidates judged to be different works	LLM	47

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.

Author Safeguards. Author lists in aggregated metadata are often incomplete. We therefore do not treat an author mismatch as a conflict unless it is supported by more reliable evidence. When a mismatch appears only in an aggregator record, we check a primary source such as arXiv or Crossref when available. Extra cited authors count as a conflict only when two independent records agree that those names are absent.

B.4 Source Resolution

Candidate Construction. 𝖱θ first groups records that refer to the same underlying work. Two records are merged when their normalized titles agree and they also share a DOI, an arXiv identifier, or the same first author. Metadata from records in the same group is consolidated: missing venue or year information can be filled from another record, and a truncated author list can be replaced by a longer compatible one.

Candidate Support. A candidate is treated as a possible source only when the retrieved evidence provides enough identity information to tie it to the citation. Concretely, we require: (1) substantial overlap between the candidate’s title and the cited title; (2) a match on either the title or the authors; (3) a record of the work itself rather than a reference-list-only entry; and (4) enough bibliographic metadata to identify the record, such as an identifier.

Once a candidate is supported, conflicting fields are treated as discrepancies to correct rather than reasons to discard the work. For example, a citation with the wrong authors can still resolve to a candidate whose title and other evidence identify the intended work.

Targeted Search and Output. If no candidate is supported, 𝖱θ performs one targeted follow-up search using the revised queries from Appendix B.2, with at most one SerpApi search and no additional page fetch. The returned records are added to the evidence and evaluated under the same support criteria.

If exactly one work is supported, 𝖱θ returns its corrected citation, with title, authors, venue, year, and identifiers taken from the retrieved records. Otherwise, it returns None.

B.5 Baseline Implementations

Direct Prompting. The backbone receives only the citation string and has no retrieval access. It directly predicts one of three outcomes: the citation is correct; the citation is incorrect but refers to an identifiable work, in which case it returns the work’s title, authors, venue, and year; or the citation is incorrect and no unique source can be identified.

CiteAudit Rerun. We rerun the released CiteAudit pipeline with the same backbone used in each comparison block, replacing its original Gemini 3 Flash extractor and judge while keeping the released prompts unchanged. The only other modification is increasing the maximum output length from 300 to 8,192 tokens to accommodate reasoning-model responses.

Web Search. This baseline uses the same source-resolution prompt as Direct Prompting, but allows the model to issue web search before answering. GPT-5.6 Luna uses OpenAI’s hosted web-search tool ⁵⁵ 5 https://developers.openai.com/api/docs/guides/tools-web-search, while OpenResearcher-30B-A3B runs in its released agent framework with the authors’ browser tools and Serper backend⁶⁶ 6 https://serper.dev/ (Li et al., 2026b).

Appendix C Evaluation Details

End-to-End Source Resolution. All systems receive only the citation string and are not told whether it is hallucinated. A prediction succeeds only if the system first identifies the citation as Fake and then either returns the correct source for a recoverable citation or returns None for an unrecoverable one. For CiteAnchor, these outputs are given by 𝖵θ and 𝖱θ, respectively. Baseline outputs are mapped to the same two decisions. A citation predicted Real therefore counts as a failure on CiteResolve.

Matching Predicted Sources to Gold Records. Returned sources are compared with the gold record field by field after normalization. Deterministic rules handle clear cases: titles must agree after normalization; author lists are matched by name while allowing initials and “et al.”; venues may match by normalized name, containment, acronym, or a preprint–published version difference; and publication years may differ by at most one year. Fields missing from the gold record are not scored.

Fields that cannot be resolved by these rules are judged by GPT-6 Sol (OpenAI, 2026b). The judge sees only the predicted and gold metadata, not the original citation or the system that produced the prediction. Each ambiguous field is judged up to three times, and the majority answer is used. As a sanity check, the judge agreed with the deterministic decision on all 100 randomly sampled fields already resolved by the rules.

Appendix D Instruction Templates

This appendix presents the instruction templates used for all LLM-based steps in our experiments: the three modules of CiteAnchor (Appendix D.1), the ablation and the baselines (Appendix D.2), and the judge used in evaluation (Appendix D.3). Each template gives the task, inputs, decision criteria, and output format of the corresponding prompt.

D.1 CiteAnchor

D.2 Ablation and Baselines

D.3 Evaluation

Appendix E More Discussion about Experiment Results

E.1 Published and Rerun CiteAudit Results

Our CiteAudit rerun with GPT-5.6 Luna scores below the published Gemini-3-Flash result (0.762 vs. 0.838 F1), mainly through lower recall (0.773 vs. 0.951) at similar precision (0.751 vs. 0.749). Since the released prompts were developed for Gemini-3-Flash, we treat the backbone-matched blocks of Table 2 as the controlled comparison and the published result as an external reference. Against this reference, CiteAnchor with GPT-5.6 Luna reaches a similar F1 (0.841) at a more conservative operating point: it flags 54 valid citations instead of 149 but misses 89 hallucinated citations instead of 23.

E.2 Resolution Outcomes by Recoverability

Table 3 evaluates two decisions in source resolution: whether to propose a source, and whether the proposed source is correct. With GPT-5.6 Luna, Web Search proposes a source for more Recoverable citations than CiteAnchor (85.3% vs. 65.9%), but a smaller share of its proposed sources match the gold source (90.2% vs. 95.2%). On Unrecoverable citations, where the correct output is to flag the citation and withhold a correction, CiteAnchor does so for 92.4% of cases and Web Search for only 31.0%. Since an incorrect correction adds a new false claim to a reference list, this combination of accurate proposals and reliable abstention gives CiteAnchor a higher overall task success (74.7% vs. 58.4%).

E.3 The Open-Weight Web Search Baseline

The low task success of Web Search with OpenResearcher-30B-A3B (5.4%) is not caused by failures of the agent framework: the model searched for 414 of the 423 citations, and all of its answers could be parsed. Instead, its corrections often repeat the errors they should fix. For Recoverable citations, it finds the right work for 129 of the 156 citations it proposes a source for, but 70 of these proposals keep the author list given in the citation, which is the erroneous field in most Recoverable citations (Table 1). For Unrecoverable citations, it proposes a source for 120 of 171, although no source can be uniquely identified. CiteAnchor instead takes the corrected source, including its author list, from retrieved records rather than from the model’s answer, allowing the same backbone to reach 70.7% task success.

注意转载授权 · 阅读原文