Hallucination

Hallucination in generative AI is the production of content that is unsupported, contradicted by an applicable source, factually wrong, internally inconsistent, or otherwise presented without an adequate basis. The term is used most often for fluent text from large language models, but related failure modes also occur in summarization, dialogue, translation, data-to-text systems, and vision-language models. Researchers do not use one universal definition. Some define hallucination by inconsistency with a supplied source, while others include conflict with established facts or failure to follow an instruction. This variation matters because a system can be faithful to a flawed source, factually correct despite adding information absent from a source, or consistently repeat the same false claim.^[1]

The National Institute of Standards and Technology uses the term "confabulation" for confidently presented false or erroneous generated content. Its definition also includes output that diverges from an input or contradicts an earlier statement in the same context. NIST treats confabulation as a risk that follows from systems generating outputs to approximate patterns in training data. That description concerns an output failure; it does not by itself establish that a model perceives, believes, or intends anything.^[2] The same caution applies to the word "hallucination": it names an observed output failure and should not be read as a literal comparison with human perception.

Hallucination is not a single model-level percentage. A measured rate depends on the task, error definition, prompt, tool access, response length, sampling procedure, abstention policy, evaluator, and date or version of the system. Results from different benchmarks are therefore not interchangeable. Reliable use requires claim-level evidence and an evaluation designed for the particular application, especially when generated content may influence legal, medical, financial, scientific, or other consequential decisions.

Definition and scope

Faithfulness, factuality, and grounding

Three related concepts help define the problem:

• Faithfulness asks whether an output is supported by the source or context it was supposed to use. A summary that changes a date from its source is unfaithful even if the changed date happens to be true elsewhere.
• Factuality asks whether a claim agrees with an appropriate account of the world. An output can be faithful to a source that is outdated or wrong and still be factually false.
• Grounding asks whether generated claims can be traced to specified evidence, such as a document, database record, image, or tool result. Grounding is always relative to the evidence provided. It does not establish that the evidence itself is authoritative, current, or correctly interpreted.

Maynez and colleagues drew a particularly useful distinction while studying abstractive summarization. They called a source contradiction an intrinsic hallucination and information that cannot be verified from the source an extrinsic hallucination. An extrinsic addition is not necessarily false; it may be accurate background knowledge. It is nevertheless unfaithful when the task requires the output to remain within the source.^[3]

Later literature on natural language processing uses broader taxonomies. One major survey of hallucination in LLMs separates factuality hallucinations from faithfulness failures, with the latter including instruction inconsistency, context inconsistency, and logical inconsistency.^[4] Other researchers reserve "hallucination" for inconsistency with an input or with information available to the model, treating factuality as a separate property. Because these schemes disagree, a study or product claim should state its reference standard instead of reporting an unexplained "hallucination rate."

What is not automatically a hallucination

Not every novel, imaginative, uncertain, or incorrect output belongs in the same category:

• A fictional story is not a hallucination merely because its events did not happen. The user requested invention.
• A creative image is not factually false merely because it depicts something nonexistent. NIST notes that non-factual visual, audio, or video generation can be the intended behavior.^[2]
• A refusal or an explicit statement of uncertainty may be unhelpful, but it is not a fabricated answer.
• A calculation, planning, or reasoning error can produce a false conclusion without fabricating a fact. Some taxonomies include such failures; others analyze them separately.
• Deception requires a separate showing of intent or strategy. A false output described as a "hallucination" does not by itself establish intent, agency, or awareness.
• Retrieval failure, stale evidence, and incorrect source interpretation are system failures that may lead to hallucinated or false output, but they can occur outside the generative model.

The boundary should be set by the task. If the instruction is "summarize only this report," any unsupported addition is relevant. If the instruction is "write a science-fiction scene," invented details are expected. If the instruction is "answer this medical question accurately," both factual correctness and appropriate uncertainty matter.

Common taxonomies

Intrinsic and extrinsic source-conditioned errors

For tasks with a defined source, the intrinsic/extrinsic distinction remains widely used:

Scroll sideways for more →

The label "extrinsic" describes verifiability from the source, not truth in the world. External evidence may later show the addition to be true or false. This is one reason source faithfulness and factuality should be scored separately.

Factual, contextual, and logical failures

In open-ended question answering there may be no supplied source. Evaluation then usually compares atomic claims with external evidence. Common error units include:

• fabricated entities, events, dates, quantities, quotations, or relationships;
• real entities combined into a relationship that the evidence does not support;
• a real source cited for a claim it does not establish;
• a response that accepts a false premise instead of correcting or qualifying it;
• contradictions between parts of the same answer;
• a conclusion that does not follow from the stated premises;
• an answer that conflicts with the user's evidence or explicit instruction.

These units overlap. A fabricated legal holding may be both factually false and misgrounded. A response may correctly name a source while misrepresenting what the source says. Evaluation at the whole-response level can conceal that mixture.

Why hallucinations occur

There is no single accepted causal account. Evidence points to failure opportunities across data collection, pretraining, post-training, retrieval, prompting, decoding, and application design. Claims about causes should identify whether they are theoretical results, correlations, controlled experiments, or mechanistic case studies.

Data and pretraining

Generative language models learn statistical regularities from large corpora. Those corpora include errors, contradictions, obsolete facts, duplicated text, source-target mismatches, and uneven coverage. Rare facts provide fewer repeated examples from which a model can learn a stable association. In conditional generation datasets, a reference answer may contain information absent from its paired input, teaching a model that unsupported additions are acceptable.^[1]

A 2025 theoretical analysis by Kalai and colleagues connects generative errors in pretraining to errors in a related binary classification problem: distinguishing valid from invalid candidate outputs. Their argument is not that every hallucination has one neural mechanism. It establishes statistical conditions under which errors can remain even if a hypothetical training set contains no false statements. The same paper argues that many common evaluations reward a guess with some chance of being correct while giving an abstention no credit, creating an incentive to answer when uncertain.^[5] This is an analysis proposed by the authors, not a universal empirical explanation of every model or task.

Post-training and evaluation incentives

Instruction tuning, preference optimization, and safety training can change whether a system answers, hedges, cites, or refuses. They do not by themselves supply a complete and current fact database. If human raters prefer confident and detailed answers, or if benchmark scoring treats "I do not know" the same as a wrong answer, optimization can favor plausible completion over calibrated abstention. Conversely, a system tuned too strongly toward refusal can reduce incorrect answers while failing to answer questions it could have handled. Accuracy, error rate, and coverage therefore need separate reporting.

Post-training may also make language more persuasive without making every underlying claim more reliable. Fluency is not a factuality test. A polished explanation, detailed citation, or step-by-step rationale can increase user trust even when the final answer or one of its premises is wrong.^[2]

Context use and generation

For source-grounded tasks, a model must select relevant details, maintain entity and quantity relationships, and preserve those relationships over a generated sequence. Errors can arise when the model relies on a superficially related pattern, gives its parametric associations more weight than supplied context, loses an earlier instruction, or continues from an already incorrect generated prefix. Sampling choices can change observed error rates in some settings, but there is no universal rule that one temperature or decoding method prevents hallucination. More deterministic output can also repeat a stable misconception.

Retrieval adds another chain of possible failures. A system may formulate the wrong search query, retrieve an irrelevant or stale document, omit the controlling source, truncate a necessary passage, or generate a claim not supported by the retrieved text. A grounded response can still be wrong when the evidence is wrong or inapplicable.

Mechanistic evidence and its limits

Interpretability research offers case-specific evidence about internal computations, not a complete theory. An Anthropic study used attribution graphs and interventions to examine selected prompts in Claude 3.5 Haiku. In its entity-recognition cases, features associated with familiar names influenced whether the model answered or indicated that it could not answer; misfires were associated with fabricated responses. The authors explicitly describe the work as selected case studies in one model, say that their replacement model is incomplete, and warn that the observed mechanisms cannot be assumed to generalize.^[6]

This result is useful as an existence proof: a model can contain mechanisms related to recognizing its knowledge boundary, and those mechanisms can fail. It does not establish a universal "hallucination circuit" or show that all models use the same process.

Hallucination across tasks and modalities

Summarization and data-to-text generation

Source-grounded summarization is a relatively clear setting because evaluators can compare a summary with its input. Errors include changed quantities, swapped entities, incorrect causal links, and unsupported additions. A summary may contain both supported and unsupported clauses in one sentence, so binary labels lose detail. Data-to-text systems face the same problem when converting tables, records, or structured fields into prose.

Source faithfulness is not the whole evaluation. A summary can omit critical qualifications while containing no direct contradiction. It can also faithfully summarize a source that is itself incorrect. Completeness, relevance, and source quality therefore remain separate dimensions.

Question answering and dialogue

Open-domain question answering is harder to score because the required evidence may be distributed across sources and may change over time. Hallucinations can appear as invented answers, false premises accepted without correction, nonexistent quotations, or citations that do not entail the associated claim. Dialogue adds conversational history: an assistant may contradict an earlier turn, attribute a statement to the wrong speaker, or prioritize a user's suggestion over better evidence.

Long answers expose more opportunities for error than single-answer tasks. Reporting only whether an entire response contains at least one false claim penalizes length and hides partial quality. Claim-level precision, coverage, and severity provide more diagnostic information.

Machine translation

In machine translation, hallucination often refers to output that is substantially detached from the source while remaining fluent in the target language. Raunak and colleagues distinguish detached hallucinations, which are fluent but severely inadequate translations, from oscillatory hallucinations containing repeated n-grams. Their controlled experiments linked these patterns to source perturbations and specific forms of noise in parallel training data, and showed that some synthetic-data procedures could amplify them.^[7] These results concern the models and experimental conditions studied; they are not a rate estimate for all translation systems.

Vision-language models

For a vision-language model, object hallucination means describing an object that is not present in the target image. The POPE study evaluated this failure as a yes-or-no object-presence problem and reported that objects common in visual instructions, or commonly co-occurring with objects actually present, were especially likely to be asserted by the models examined.^[8] Object presence is only one part of multimodal faithfulness. Attributes, counts, spatial relationships, text in an image, and event descriptions can fail separately.

This use of "hallucination" should not be extended indiscriminately to image generation. A caption that says an absent bottle is visible conflicts with an input image. A requested illustration of an imaginary bottle does not.

Documented consequences

Scholarly citations

Fabricated references are a concrete and testable failure. Walters and Wilder used GPT-3.5 and GPT-4 in the first week of April 2023 to produce 84 short literature reviews across 42 topics, yielding 636 bibliography entries. In that experiment, 55% of GPT-3.5 citations and 18% of GPT-4 citations referred to works the researchers could not locate. Among citations to real works, 43% and 24%, respectively, contained substantive bibliographic errors.^[9] These figures describe specific model versions, prompts, subjects, and search procedures from 2023. They should not be presented as current rates for all scholarly use.

The experiment also illustrates why a plausible title, author list, journal name, or DOI-like string is not evidence that a source exists. Bibliographic verification requires locating the work in an authoritative catalog, publisher site, or index and checking that it supports the associated claim.

Legal research

In Mata v. Avianca, a United States federal court found that attorneys submitted nonexistent judicial opinions with fake quotations and citations created by ChatGPT, then continued to rely on them after the court questioned their existence. The June 22, 2023 order imposed a joint $5,000 penalty and required notices to the client and judges falsely named as authors of the fake opinions.^[10] The sanction was based on the respondents' conduct and legal duties, not on a general judicial finding about every use of generative AI.

Domain retrieval does not remove the need to verify. A preregistered study later published in the Journal of Empirical Legal Studies tested leading proprietary legal research assistants. More than one in six evaluation queries caused each of two products to return misleading or false information under the study's definition, while a third product did so on about one-third of queries. The systems differed substantially in refusal and accuracy, and their errors included irrelevant retrieval, inapplicable authority, reasoning failures, and false source support.^[11] These are product-specific measurements from a defined query set and point in time.

Health information

The World Health Organization's 2024 guidance on large multimodal models for health identifies documented risks of false, inaccurate, biased, or incomplete statements. It also warns about automation bias, in which clinicians or patients overlook errors or improperly delegate difficult choices. WHO recommends well-defined tasks, stakeholder participation, regulatory assessment, independent post-release auditing, and impact assessment for large-scale deployments.^[12] Those controls address the surrounding system and workflow, not only the base model.

Compounding and automation

A single incorrect claim can be copied into reports, search indexes, decisions, or later prompts. When users see a fluent answer, a citation, or an explanation, they may give it more weight than its evidence warrants. High-volume generation can scale this problem, but severity depends on the use. A wrong film date in casual conversation is unlike an invented contraindication in a clinical workflow. Risk assessment should therefore combine frequency with consequence, detectability, reversibility, and the availability of expert review.

Detection and measurement

No detector establishes truth in every domain. Detection methods differ in what evidence they require and which failure mode they can see.

Reference-based methods

When an authoritative source is available, evaluators can split a response into claims and test whether each claim is entailed, contradicted, or not addressed by the source. Methods include lexical comparison, natural language inference, question-answering checks, database lookup, and human review. A source-based score should distinguish contradiction from absence of support.

FActScore operationalizes factual precision for long-form generation by decomposing text into atomic facts and calculating the proportion supported by a designated knowledge source.^[13] That design reveals mixed-quality passages better than a single response label. It measures precision rather than factual recall, so a terse or frequently abstaining system can earn a high score while omitting useful information.

Reference-based evaluation inherits the limitations of its evidence. Retrieval may miss a source; a database may be incomplete; a document may be outdated; and an entailment model may misunderstand negation, time, jurisdiction, or qualification. The evidence set and its date should be recorded.

Reference-free and consistency methods

SelfCheckGPT compares a response with multiple stochastic samples from the same black-box model. The premise is that a model with stable knowledge will tend to produce semantically consistent facts, whereas arbitrary fabrication will vary across samples.^[14] This can flag uncertainty without an external database, but it can miss a misconception that the model repeats consistently. It also increases inference cost and depends on sampling and comparison choices.

Semantic entropy groups sampled answers by meaning before measuring uncertainty, so wording differences such as "Paris" and "the capital is Paris" do not count as different answers. Farquhar and colleagues explicitly target confabulations: arbitrary, incorrect generations that vary with sampling. Their method does not guarantee factuality and is not designed to catch consistent learned falsehoods, deception, or systematic reasoning errors.^[15]

Other reference-free signals include token probability, stated confidence, cross-question consistency, and model-based critique. Low confidence can accompany a correct answer, while high confidence can accompany a stable error. These signals are triage tools rather than verdicts.

Human evaluation and automatic judges

Human review remains necessary for nuanced or consequential claims, but it is not automatically reliable. A 2025 survey examined 64 studies with human hallucination evaluation published from 2019 through 2024. It found substantial variation in definitions and frequent under-reporting of guidelines, interface design, inter-annotator agreement, and annotator demographics.^[16] Evaluators also need relevant subject expertise when truth depends on medicine, law, science, or another specialized field.

Automatic metrics can be useful at scale, but their scores are not interchangeable. A 2025 evaluation of six metric families across four datasets, 37 language models, and five decoding methods found limited agreement among metrics and inconsistent alignment with human judgments. LLM-based judging performed best overall in that study, but the authors still identified generalization and evaluation limitations.^[17] A model judge may share errors with the system it evaluates or prefer outputs from a related model family. High-stakes evaluation should include blinded human adjudication and a documented rubric.

Benchmarks

Benchmarks isolate different slices of the problem:

Scroll sideways for more →

Leaderboard values should be accompanied by the model snapshot, prompt, decoding configuration, tool permissions, judge version, sample count, confidence interval when available, and evaluation date. Comparing two systems on different benchmarks, or comparing a browsing system with a closed-book model, does not establish which one hallucinates less in actual use.

Mitigation

Mitigation is a layered system-design problem. No method below is a general guarantee.

Retrieval and browsing

Retrieval-augmented generation supplies documents at inference time instead of relying only on parametric memory. In the original RAG experiments, retrieved Wikipedia passages improved several knowledge-intensive tasks, and human evaluators preferred RAG over a parametric baseline for factuality and specificity in the studied question-generation setting.^[23] Retrieval can also provide provenance and allow an index to be updated without retraining the generator.

The benefit depends on retrieving the right evidence and using it faithfully. In law, medicine, or policy, applicability may depend on jurisdiction, date, population, or procedural context. Systems should expose the actual passage used, not just a homepage or document title, and preserve metadata needed to judge authority.

Citations and claim traceability

A citation makes a claim auditable only when the cited passage exists and supports that claim. The ALCE benchmark evaluates answer correctness, citation correctness, and citation completeness. In its ELI5 experiments, even the strongest systems lacked complete citation support for about half of outputs.^[24] The result demonstrates that adding citation markers does not ensure that every claim is supported.

Practical safeguards include requiring citations at the sentence or atomic-claim level, blocking citations that were not returned by the retrieval system, validating identifiers against authoritative records, and checking entailment between each claim and its cited passage. A verifier should also look for uncited factual clauses.

Independent verification and revision

Verification works better when it is separated from drafting. Chain-of-Verification first creates a draft, plans questions that would test its factual claims, answers those questions independently, and then revises the draft. The method reduced errors on the list, closed-book question-answering, and long-form tasks studied, but the authors state that it did not remove hallucinations completely.^[25]

Independence matters. A verifier shown only the original answer may repeat or rationalize the same mistake. Stronger workflows retrieve fresh evidence for each material claim, use a different model or deterministic validator where appropriate, and escalate conflicts to a person with domain expertise.

Uncertainty and abstention

Systems should be allowed to say that available evidence is insufficient. Evaluation should report at least:

• coverage, the proportion of requests answered;
• accuracy among answered requests;
• overall incorrect-answer rate;
• calibration, the relationship between stated confidence and observed correctness;
• selective risk, the error rate after low-confidence responses are withheld.

An abstention policy must be tested for both false answers and false refusals. A system that never answers avoids hallucination but has no utility. A system that always guesses maximizes coverage while exposing users to unsupported claims. The acceptable balance depends on consequence and the availability of fallback review.

Task constraints and structured interfaces

Applications can reduce opportunities for unsupported free-form output by narrowing the task. Examples include extracting specified fields, selecting from validated entities, requiring a quotation span before a summary claim, constraining output to a schema, computing numbers with deterministic code, and resolving identifiers through an authoritative knowledge graph. Constraints do not prove that upstream data or business rules are correct, but they make failure easier to detect.

Prompt instructions can help define the evidence boundary, desired uncertainty behavior, and citation format. They should be treated as one control, not a security or factuality guarantee. A system can ignore, forget, or misinterpret a prompt, especially over long interactions.

Training, evaluation, and monitoring

Data curation can reduce duplicated, contradictory, or source-divergent examples. Fine-tuning can reward source faithfulness, correction of false premises, appropriate refusal, and calibrated uncertainty. Evaluations should give partial credit for justified abstention rather than always rewarding guesses.^[5]

Production monitoring should use representative traffic, preserve model and retrieval versions, sample outputs for expert review, track failures by severity and subgroup, and retest after any model, prompt, index, or policy change. An aggregate score can hide a rare but serious failure. Red-team cases should include unanswerable questions, false premises, stale documents, conflicting sources, long contexts, adversarial instructions, and requests outside the system's declared scope.

Operational safeguards

The following controls are appropriate when output may affect consequential decisions:

Scroll sideways for more →

For health deployments, these controls align with WHO's emphasis on well-defined tasks, stakeholder participation, oversight, independent auditing, and impact assessment.^[12] For legal and scholarly work, every proposition and citation remains the responsibility of the person or organization using it.

Limits and open questions

Research has not produced a universal detector or a model-independent rate of hallucination. Some errors are arbitrary and vary across samples; others are stable misconceptions. Some are visible against a supplied document; others require current specialist knowledge. A system may improve on one benchmark by refusing more often, producing shorter answers, or learning the benchmark distribution without becoming more reliable in deployment.

Open questions include how to evaluate claims whose truth changes over time, how to measure completeness without rewarding unsupported detail, how to communicate calibrated uncertainty in natural language, how to audit proprietary retrieval pipelines, and how to detect consistent errors that self-consistency methods miss. Terminology is also unsettled. Clear operational definitions and transparent evaluation protocols are more useful than a claim that a system is simply "hallucination-free."

See also

• AI safety
• Prompt Engineering
• Artificial Intelligence
• Language Model

References

1. ^Ziwei Ji et al. "Survey of Hallucination in Natural Language Generation." ACM Computing Surveys, 2023. doi.org/...3571730
2. ^National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1), 2024. doi.org/...NIST.AI.600-1
3. ^Joshua Maynez et al. "On Faithfulness and Factuality in Abstractive Summarization." Proceedings of ACL 2020, 2020. aclanthology.org/2020.acl-main.173
4. ^Lei Huang et al. "A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions." ACM Transactions on Information Systems, 2025. doi.org/...3703155
5. ^Adam Tauman Kalai et al. "Why Language Models Hallucinate." 2025. openai.com/...why-language-models-hallucinate
6. ^Jack Lindsey et al. "On the Biology of a Large Language Model." Anthropic, 2025. transformer-circuits.pub/...biology
7. ^Vikas Raunak, Arul Menezes, and Marcin Junczys-Dowmunt. "The Curious Case of Hallucinations in Neural Machine Translation." Proceedings of NAACL 2021, 2021. aclanthology.org/2021.naacl-main.92
8. ^Yifan Li et al. "Evaluating Object Hallucination in Large Vision-Language Models." Proceedings of EMNLP 2023, 2023. aclanthology.org/2023.emnlp-main.20
9. ^William H. Walters and Esther Isabelle Wilder. "Fabrication and Errors in the Bibliographic Citations Generated by ChatGPT." Scientific Reports, 2023. doi.org/...s41598-023-41032-5
10. ^Mata v. Avianca, Inc., No. 22-cv-1461, Opinion and Order on Sanctions, U.S. District Court for the Southern District of New York, June 22, 2023. courtlistener.com/...mata-v-avianca-inc
11. ^Varun Magesh et al. "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools." Journal of Empirical Legal Studies, 2025. doi.org/...jels.12413
12. ^World Health Organization. "WHO Releases AI Ethics and Governance Guidance for Large Multi-Modal Models." 2024. who.int/...e-guidance-for-large-multi-modal-models
13. ^Sewon Min et al. "FActScore: Fine-Grained Atomic Evaluation of Factual Precision in Long Form Text Generation." Proceedings of EMNLP 2023, 2023. aclanthology.org/2023.emnlp-main.741
14. ^Potsawee Manakul, Adian Liusie, and Mark J. F. Gales. "SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models." Proceedings of EMNLP 2023, 2023. aclanthology.org/2023.emnlp-main.557
15. ^Sebastian Farquhar et al. "Detecting Hallucinations in Large Language Models Using Semantic Entropy." Nature, 2024. doi.org/...s41586-024-07421-0
16. ^Patricia Schmidtova et al. "Do My Eyes Deceive Me? A Survey of Human Evaluations of Hallucinations in NLG." Proceedings of INLG 2025, 2025. aclanthology.org/2025.inlg-main.4
17. ^Atharva Kulkarni et al. "Evaluating Evaluation Metrics: The Mirage of Hallucination Detection." Findings of EMNLP 2025, 2025. aclanthology.org/2025.findings-emnlp.1035
18. ^Stephanie Lin, Jacob Hilton, and Owain Evans. "TruthfulQA: Measuring How Models Mimic Human Falsehoods." Proceedings of ACL 2022, 2022. aclanthology.org/2022.acl-long.229
19. ^Junyi Li et al. "HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models." Proceedings of EMNLP 2023, 2023. aclanthology.org/2023.emnlp-main.397
20. ^Jason Wei et al. "Introducing SimpleQA." OpenAI, 2024. openai.com/...introducing-simpleqa
21. ^FACTS Team. "FACTS Grounding: A New Benchmark for Evaluating the Factuality of Large Language Models." Google DeepMind, 2024. deepmind.google/...uality-of-large-language-models
22. ^Yejin Bang et al. "HalluLens: LLM Hallucination Benchmark." Proceedings of ACL 2025, 2025. aclanthology.org/2025.acl-long.1176
23. ^Patrick Lewis et al. "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." Advances in Neural Information Processing Systems 33, 2020. papers.nips.cc/...05f780e1bc26945df7481e5-Abstract
24. ^Tianyu Gao et al. "Enabling Large Language Models to Generate Text with Citations." Proceedings of EMNLP 2023, 2023. aclanthology.org/2023.emnlp-main.398
25. ^Shehzaad Dhuliawala et al. "Chain-of-Verification Reduces Hallucination in Large Language Models." Findings of ACL 2024, 2024. aclanthology.org/2024.findings-acl.212