Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents

Abstract: LLMs and deep research agents supply citation URLs to support their claims, yet the reliability of these citations has not been systematically measured. We address six research questions about citation URL validity using 10 models and agents on DRBench (53,090 URLs) and 3 models on ExpertQA (168,021 URLs across 32 academic fields). We find that 3--13\% of citation URLs are hallucinated -- they have no record in the Wayback Machine and likely never existed -- while 5--18\% are non-resolving overall. Deep research agents generate substantially more citations per query than search-augmented LLMs but hallucinate URLs at higher rates. Domain effects are pronounced: non-resolving rates range from 5.4\% (Business) to 11.4\% (Theology), with per-model effects even larger. Decomposing failures reveals that some models fabricate every non-resolving URL, while others show substantial link-rot fractions indicating genuine retrieval. As a solution, we release urlhealth, an open-source tool for URL liveness checking and stale-vs-hallucinated classification using the Wayback Machine. In agentic self-correction experiments, models equipped with urlhealth reduce non-resolving citation URLs by $6\textrm{--}79\times$ to under 1\%, though effectiveness depends on the model's tool-use competence. The tool and all data are publicly available. Our characterization findings, failure taxonomy, and open-source tooling establish that citation URL validity is both measurable at scale and correctable in practice.

• The paper quantifies citation URL failures across models, with non-resolving rates up to 18% and hallucinated citations reaching 13%.
• It employs HTTP liveness checks and Wayback Machine verification to differentiate between fabricated URLs and genuine link rot.
• Tool-based mitigation via urlhealth reduces hallucination errors dramatically, achieving final rates below 1% through iterative self-correction.

Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents: A Technical Synthesis

Introduction and Problem Scope

Reference hallucination—the generation of plausible yet nonexistent citations—remains a pervasive failure mode in LLMs and agentic research architectures. These entities are increasingly deployed in retrieval-augmented generation (RAG) and "deep research" settings, providing inline citations users may treat as ground truth. This paper delivers the first large-scale, systematic analysis of citation URL validity in leading commercial LLMs and deep research agents, providing measurements and interventions across both model and domain axes. Six research questions structure the empirical inquiry: prevalence of citation URL failures, system-type effects, domain variation, failure-type decomposition (fabrication vs. link rot), volume-quality trade-offs, and post-hoc mitigation efficacy.

Datasets and Methodology

The study evaluates 10 models across DRBench (53,090 URLs, cross-model perspective) and three models across ExpertQA (168,021 URLs, domain-stratified). DRBench allows comparison between providers (Google, OpenAI, Anthropic) and system categories (deep research agents vs. search-augmented LLMs). ExpertQA addresses field-level variation across 32 academic domains.

Citation URLs are extracted, canonicalized, and probed for liveness via HTTP HEAD/GET requests. Non-resolving URLs (4xx, 5xx, connection failures) are further checked in the Wayback Machine to sub-classify hallucinated (never-archived; likely fictional) versus stale (was archived; genuine link rot) URLs. These distinctions reveal both generative and retrieval-based failure phenotypes.

Empirical Results

Prevalence and Model Effects

Non-resolving URL rates span 5–18% across commercial LLMs/agents, with hallucinated URLs a 3–13% subset.

Figure 1: Non-resolving URL rates for DRBench models, grouped by provider, partitioned into hallucinated (red) and stale (orange) URLs.

Deep research agents (e.g., Gemini-2.5-Pro Deep Research, OpenAI Deep Research) generate more citations per query but incur higher hallucination rates (mean 10.7%) than search-augmented LLMs (mean 4.8%). The link between retrieval architecture and hallucination dwarfs the effect of raw citation throughput; multi-step synthesis exacerbates fabrication, potentially via compositional error propagation in candidate generation. Some models (notably certain OpenAI LLMs) fabricate every non-resolving URL, exhibiting no trace of legitimate but decayed links, indicating the absence of retrieval grounding.

Domain and Subfield Variation

Non-resolving rates differ by a factor of ~2x across fields in ExpertQA, from 5.4% (Business) to 11.4% (Theology). Fine-grained analysis within Healthcare/Medicine exposes a 14.8–21.4% interval among subfields.

Figure 2: Non-resolving URL rates by academic field and model for ExpertQA.

Figure 3: Non-resolving URL rates by subfield within Healthcare/Medicine demonstrate large intra-domain variability.

Medical and scientific fields—domains with compounding real-world risk upon erroneous citations—systematically display elevated rates, with LLM-dependent relative ordering.

Hallucination versus Link Rot

The hallucinated/stale decomposition is clarifying. Google and Anthropic search-augmented models present a meaningful stale-component, indicating true retrieval with subsequent link disappearance. In sharp contrast, OpenAI’s smaller LLMs uniformly hallucinate non-resolving citations, generating URLs unanchored to retrievable content. This dichotomy signals whether the failure is fundamentally generative (fabrication) or infrastructural (dataset staleness).

Citation Volume versus Reliability

Higher citation volume correlates inversely with reliability; models producing more citations per query (e.g., GPT-5.1 at 46.4, Gemini at 10.7) accrue higher per-citation error rates. This tendency holds within and across system classes, refuting hypotheses that increased citation density acts as a check on hallucination.

Tooling and Mitigation: urlhealth and Self-Correction

The authors introduce urlhealth, an open-source tool implementing URL liveness checks with classification (live, dead, hallucinated, unknown). This tool is deeply evaluated in an agentic self-correction framework where models iteratively revise outputs conditioned on external verification feedback.

Figure 4: Distribution of urlhealth correction rounds per question shows distinct self-correction behavior across models; all converge to negligible hallucination rates in final output.

Post-mitigation with urlhealth achieves an order-of-magnitude reduction in non-resolving citations (6–79x reduction). Final hallucinated URL rates fall below 1%. Tool-use competence, however, emerges as a new bottleneck: smaller models (e.g., GPT-5-nano) repeatedly propose flagged URLs, failing agentically at exploitation of verification signals despite infrastructure support.

Related Work and Implications

Previous studies quantified reference hallucination in non-retrieval LLMs, with GPT-3.5 fabricating 55% of citations, GPT-4 fabricating 18% [walters2023fabrication], and strong popularity bias in generated references [algaba_large_2024, algaba_how_2025]. Domain effects have also been reported [mugaanyi2024doi, dahl2024legal, linardon2025mental]. Recent large-scale audits (GhostCite, CiteAudit) exposed persistent hallucination and weak reviewer/copyeditor verification processes at scale [xu2026ghostcite, yuan2026citeaudit]. Most concurrent RAG benchmarks treat source existence as axiomatic; this work demonstrates that nonexistence/fabrication remains an unresolved failure axis even under retrieval augmentation.

On the mitigation front, urlhealth complements approaches like SemanticCite [haan_semanticcite_2025] and CiteGuard [choi_citeguard_2026], which focus more on source-claim matching rather than source existence. The findings stress a required shift from implicit trust in RAG to explicit, tool-mediated pipeline verification, especially in peer-reviewed and high-stakes outputs.

Practical and Theoretical Implications

Practically, URL hallucination undermines user trust in LLM-cited evidence and is now detectable and correctable at production scale. Future LLM/agent deployments should restrict generation-time emission to actually visited URLs, moving models toward robustly grounded text generation. Observations about agentic tool-use failures suggest additional research in planning and reasoning for tool-aware LLMs will be critical for closing the reliability gap in bibliographic and knowledge grounding.

Theoretically, the existence of non-resolving but parameterically plausible URLs calls for a revised taxonomy of LLM hallucination in RAG-mediated regimes: even with retrieval, fabrication persists, and volume-based heuristics for trustworthiness are invalid. Structural fidelity of generated knowledge, as studied in network-centric analyses [boudourides2026structural], cannot be decoupled from local source existence. Finally, the persistence of hallucination in retrieval-augmented agents suggests that LLM training and decoding should strongly incentivize verifiable grounding not just in content, but also in referenced sources.

Conclusion

Systematic, large-scale evaluation reveals that hallucinated citation URLs (3–13%) and overall non-resolving rates (5–18%) persist in state-of-the-art LLMs and deep research agents, irrespective of retrieval augmentation. Deep research agents exacerbate fabrication despite increased citation throughput. Domain and model effects are both substantial, and failure analysis (stale versus hallucinated) surfaces critical structural differences in grounding paradigms. Post-hoc verification with urlhealth, when coupled with competent agentic tool use, robustly eliminates reference hallucination. These findings demand both infrastructural and model-centric approaches to enforce source existence, with direct implications for scientific publishing, regulatory compliance, and safety in AI-mediated research workflows.

---

References

• Rao, D. et al. "Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents" (2604.03173).
• Algaba, A. et al. "LLMs Reflect Human Citation Patterns with a Heightened Citation Bias" (Algaba et al., 2024).
• Walters, W. H. & Wilder, E. I. "Fabrication and errors in the bibliographic citations generated by ChatGPT," Scientific Reports (2023).
• Xu, Z. et al., "GhostCite: A Large-Scale Analysis of Citation Validity in the Age of LLMs" (Xu et al., 6 Feb 2026).
• Yuan, Z. et al., "CiteAudit: You Cited It, But Did You Read It? A Benchmark for Verifying Scientific References in the LLM Era" (Yuan et al., 26 Feb 2026).
• Haan, S., "SemanticCite: Citation Verification with AI-Powered Full-Text Analysis and Evidence-Based Reasoning" (Haan, 20 Nov 2025).
• Choi, Y. M. et al., "CiteGuard: Faithful Citation Attribution for LLMs via Retrieval-Augmented Validation" (Choi et al., 15 Oct 2025).
• Boudourides, M., "Structural Hallucination in LLMs: A Network-Based Evaluation of Knowledge Organization and Citation Integrity" (Boudourides, 2 Mar 2026).