Survey of Hallucination in Natural Language Generation

• Transformer-based NLG has expanded across summarization, dialogue generation, generative question answering, data-to-text generation, and machine translation [44].Source: §1 · p. 3
• Likelihood-maximization approaches can produce bland, incoherent, or repetitive degeneration, while other models generate nonsensical or source-unfaithful text termed hallucination [215].Source: §1 · p. 3
• Hallucinations threaten real-world performance, safety, and privacy, including potentially dangerous medical or translation outputs and recovery of sensitive training-corpus information [26].Source: §1 · p. 3
• The survey addresses a gap left by task-specific reviews [104] through a unified analysis of hallucination research across major unimodal NLG tasks and selected visual-language settings [3] [17].It emphasizes similarities and differences in concepts, metrics, and mitigation methods while focusing mainly on textual input sources.Source: §1 · p. 3
• In computer vision, “hallucination” may describe beneficial superresolution, inpainting, or image synthesis, whereas related errors in image captioning and object detection involve nonexistent or mislocalized objects [17].Source: §1 · p. 3
• The survey proceeds from definitions, contributors, metrics, and mitigation methods to task-specific studies, visual-language generation, LLM hallucinations, and a concluding synthesis.Sections 2–6 cover general foundations; Sections 7–12 cover downstream tasks; Section 13 covers LLMs; Section 14 concludes the survey.Source: §1 · p. 4

2 DEFINITIONS

• NLG hallucination is generated content that is nonsensical or unfaithful to provided source content, often appearing fluent and grounded despite being difficult to detect.The terminology is motivated by psychological hallucination: an unreal perception that feels real.Source: §2 · p. 4

2.1 Categorization

• Intrinsic hallucinations contradict source content, whereas extrinsic hallucinations cannot be verified from it; extrinsic information may be correct but remains risky because it is unverifiable [60] [104].For example, stating that an Ebola vaccine was approved in 2021 contradicts a 2019 source, while unmentioned COVID-19 vaccine trials are unverifiable.Source: §2.1 · p. 4

2.2 Task Comparison

• Hallucination definitions and tolerance differ by task: faithfulness is essential in summarization and data-to-text, while dialogue allows more tolerance to support user engagement.Sources are input text for summarization, non-linguistic data for data-to-text, and dialogue history or external knowledge for dialogue [109].Source: §2.2 · p. 5
• Generative question answering lacks a standard hallucination definition and mainly studies intrinsic hallucination, while machine translation commonly treats outputs disconnected from the source as hallucinations [215].These task-specific differences are documented further in Section 11.Source: §2.2 · p. 5

2.3 Terminology Clarification

• Faithfulness means consistency with the provided source, whereas factuality concerns actuality or fact, so the two concepts depend on whether the fact is source knowledge or world knowledge [50].The survey adopts Maynez et al.’s distinction, treating world knowledge as the basis of factuality, and includes work aimed at maximizing faithfulness.Source: §2.3 · p. 5

3.1 Hallucination from Data

• Source-reference divergence is the main data-related cause of hallucination, encouraging models trained on mismatched pairs to generate text unsupported by the source.Divergence can result from heuristic data collection or from task characteristics that inherently permit weak source-target alignment.Source: §3.1 · p. 5
• 62% of WIKIBIO first sentences contain information absent from their corresponding infoboxes, illustrating how heuristic source-target mismatches can cause hallucination [45].WIKIBIO pairs Wikipedia infoboxes with article first sentences, which are not necessarily equivalent in information content.Source: §3.1 · p. 7
• Duplicate training examples can bias models toward repeating memorized phrases, while dialogue tasks may inherently permit extrinsic hallucinations to improve engagingness and diversity [134].Open-domain dialogue may include chit-chat, subjective responses, or relevant facts absent from the input, history, or knowledge source.Source: §3.1 · p. 8

3.2 Hallucination from Training and Inference

• Hallucination persists even with little dataset divergence because neural-model training and modeling choices introduce additional errors [215].Imperfect encoders may learn defective representations or wrong correlations, causing generations to diverge from the input [6] [251].Source: §3.2 · p. 8
• Wrong decoder attention can mix facts between similar entities, while top-k sampling increases hallucination and exposure bias compounds errors as generated sequences lengthen [251] [60] [13].Training with teacher-forced ground-truth prefixes differs from inference on self-generated history, and this discrepancy can produce increasingly erroneous generations.Source: §3.2 · p. 8
• Parametric knowledge can override provided input, causing excess information in outputs, while models may still recognize that some generations are hallucinated [21] [113] [174].Pretraining memorizes knowledge in model parameters, improving generalizability and coverage but also contributing to hallucination [169].Source: §3.2 · pp. 8–9

4.1 Statistical Metric

• Statistical metrics use lexical overlap or entailment between generated, source, and reference texts, but lexical matching misses syntactic and semantic variation [45].PARENT [45] combines source-table and target-text entailment through an F1-score, while source-only variants address one-to-many outputs.Source: §4.1 · pp. 9–10

4.2 Model-based Metric

• Model-based metrics capture complex content mismatches, yet errors in their neural components can propagate and distort hallucination estimates.Information-extraction metrics verify source and generated relation tuples, such as differing birth years, while QA-based metrics compare answers generated from the output against answers from the source.Source: §4.2 · p. 10, §4.2.1 · p. 10, §4.2.2 · p. 10

4.2.1 Information Extraction (IE)-based.

• IE-based metrics extract relational facts from generated and reference texts, then identify hallucinations through mismatched tuples, but remain vulnerable to extraction errors.This representation excludes non-verifiable words and focuses verification on extracted knowledge-bearing relations.Source: §4.2.1 · p. 10

4.2.2 QA-based.

• QA-based metrics generate questions from the output, answer them using the source, and score hallucination from corresponding-answer similarity across several NLG tasks [54].Errors from either the question-generation or question-answering model can propagate into the metric.Source: §4.2.2 · p. 10

4.2.3 Natural Language Inference (NLI) Metrics.

• NLI-based metrics estimate faithfulness from source-to-generation entailment, contradiction, and neutrality, offering greater robustness to lexical variability than token matching.However, off-the-shelf NLI models transfer poorly to abstractive summarization, sentence-level entailment cannot localize errors, and task-specific datasets or dependency-level entailment address these limitations [160] [226].Source: §4.2.3 · pp. 10–11, §4.2.4 · p. 11

4.2.4 Faithfulness Classification Metrics.

• Faithfulness-specific classification datasets better reflect hallucination than generic NLI labels because neutral or entailed statements can still contain unsupported information.For example, adding “U.S.” to “Putin is president” introduces unsupported content even if the resulting hypothesis is neutral or entailed under NLI.Source: §4.2.4 · p. 11

4.2.5 LM-based Metrics.

• LM-based metrics classify tokens as hallucinatory when an unconditional language model assigns lower loss than a source-conditioned model, using the hallucinated-token ratio as the degree.The comparison is performed during forced-path decoding against ground-truth targets [251].Source: §4.2.5 · p. 11

4.3 Human Evaluation

• Because automatic hallucination evaluation remains imperfect, human evaluation commonly uses scoring or comparison against baselines and references [226].Human studies use terms including faithfulness, factual consistency, fidelity, factuality, and hallucination, and may separately assess intrinsic and extrinsic hallucinations [32].Source: §4.3 · p. 11

5.1.1 Building a Faithful Dataset.

• Faithful datasets are built by writing or revising targets, trimming unsupported phrases, resolving context dependence, modifying syntax, or labeling model-generated outputs for hallucinations.Existing parallel corpora can also be filtered or corrected for irrelevant or contradictory information, especially when the original noise is low or moderate.Source: §5.1.1 · p. 12, §5.1.2 · p. 12

5.1.2 Cleaning Data Automatically.

• Automatic data cleaning filters hallucinated source-reference pairs, corrects structured inputs using meaning representations extracted from references, and augments inputs with entities, relations, operations, synthetic data, retrieved knowledge, or similar samples.Instance-level filtering may lose word-level divergence signals, whereas input correction preserves the dataset and improves semantic consistency between inputs and outputs.Source: §5.1.2 · p. 12, §5.1.3 · p. 12

5.1.3 Information Augmentation.

• Information augmentation strengthens input-output alignment, but semantic gaps and format discrepancies between original sources and added structured or external information create new challenges.Examples include ambiguous utterances paired with distinct meaning representations and structured knowledge graphs paired with natural language.Source: §5.1.3 · p. 12

5.2.1 Architecture.

• Architecture-based methods improve source representation and faithfulness through dual encoders, source-conditioned or sparse attention, inductive attention, specialized decoders, and constrained decoding.Decoder constraints and reduced sampling randomness can increase faithful tokens, but may reduce fluency or diversity.Source: §5.2.1 · p. 13

5.2.2 Training.

• Training and generation methods control hallucination through planning or sketching, reinforcement-learning rewards, multi-task objectives, controllable generation, regularization, loss reconstruction, and post-processing correction.These methods trade off faithfulness against diversity, convergence difficulty, task-selection and optimization challenges, annotation needs, or grammaticality; generate-then-refine correction is particularly competitive for noisy datasets.Source: §5.2.2 · pp. 13–14
• Post-processing correction can use small amounts of automatically generated training data to refine outputs from strong fluency-oriented models, although correction may produce ungrammatical text.The generate-then-refine strategy is presented as a competitive choice when noisy datasets contain many hallucinated references.Source: §5.2.3 · p. 14

6.1 Future Directions in Metrics Design

• Fine-grained metrics should locate and distinguish intrinsic from extrinsic hallucinatory substrings rather than treating them as one unified score.Automatic substring detection and categorization would improve metric insight and explainability, especially for metrics such as QA-based methods that cannot localize individual errors.Source: §6.1 · p. 15
• Automatic fact-checking requires both broader, trustworthy evidence retrieval from world knowledge and verification models robust to adversarial, negation, numerical, and comparative changes.Current evidence sources range from limited Wikipedia coverage to the whole web, whose trustworthiness remains difficult to ensure; verification models perform well with correct evidence but remain vulnerable to small wording changes.Source: §6.1 · p. 15
• Future metrics should generalize across tasks and domains, support unified benchmarks, and model human sensitivity to salient information such as named entities and numerals.Open-source platforms and automatic check-worthy detection are proposed to standardize evaluation and improve correlation with human judgment.Source: §6.1 · pp. 15–16

6.2 Future Directions in Mitigation Methods

• Mitigation research should develop robust cross-task preprocessing and explicitly model numerals, whose errors in dates, quantities, and scalars can be important or fatal.Existing preprocessing can introduce grammatical or semantic changes, while proposed numeral approaches use commonsense knowledge or rerank summaries using quantity-entity verification scores.Source: §6.2 · p. 16
• Mitigation methods should separately address overlooked extrinsic hallucinations, long-input and self-contradictory long-output settings, and reasoning over source facts and quantities.Suggested directions include using fact-checking methods, Longformer-based encoding, reducing contradictions with dialogue history, and adding forward or backward numerical and textual reasoning.Source: §6.2 · pp. 16–17
• Controllable mitigation should balance faithfulness and diversity because acceptable hallucination levels differ between factual chit-chat and abstractive summarization.The survey notes that the research community has not reached agreement on whether factual hallucinations are desirable in abstractive summaries.Source: §6.2 · p. 17

7 HALLUCINATION IN ABSTRACTIVE SUMMARIZATION

• 25% of summaries from state-of-the-art abstractive summarization models contain hallucinated content, and high ROUGE scores can coexist with substantial hallucination.Hallucination motivates evaluation methods that assess faithfulness rather than relying only on surface-overlap metrics.Source: §7 · p. 17

7.1 Hallucination Definition in Abstractive Summarization

• A summary is hallucinated when it contains spans unsupported by the document; intrinsic hallucination contradicts the source, whereas extrinsic hallucination is unverifiable from it.Pagnoni et al. further classify factual errors as semantic frame, discourse, and content verifiability errors.Source: §7.1 · p. 17

7.2 Hallucination Metrics in Abstractive Summarization

• Hallucination metrics are mainly model-based and divided into unsupervised and semi-supervised categories, but they generally evaluate intrinsic and extrinsic hallucinations together because automatic separation is difficult.Unsupervised metrics include information extraction, natural language inference, and question answering approaches.Source: §7.2 · pp. 17–18, §7.2.1 · p. 18

7.2.1 Unsupervised Metrics.

• QA-based metrics such as FEQA [54], QAGS, and QuestEval generate questions, compare source and summary answers, and correlate better with human faithfulness judgments than baseline metrics.Their effectiveness depends on the generated questions, motivating meta-evaluation of QA metrics.Source: §7.2.1 · p. 18

7.2.2 Semi-Supervised Metrics.

• FactCC outperforms classifiers trained on MNLI or FEVER, while Zhou et al. report higher human-consistency correlations after fine-tuning on synthetic hallucinations.These semi-supervised methods learn from task-specific synthetic data to classify factual consistency or identify unfaithful spans.Source: §7.2.2 · pp. 18–19

7.3.1 Architecture Method.

• Mitigation methods modify encoders, decoders, or both using fact graphs, external knowledge, focused attention, dual attention, and entailment-aware multi-task learning [6] [25].These approaches aim to condition generation more strongly on source facts and improve summary faithfulness.Source: §7.3.1 · p. 19

7.3.2 Training Method.

• Training-based mitigation contrasts reference summaries with automatically generated hallucinations, while CONFIT improves factual consistency and overall quality in dialogue summarization [23] [245].Post-editing methods then correct unsupported spans, verify quantities, or replace generated entities with source-supported ones [50] [22] [32].Source: §7.3.2 · p. 19, §7.3.3 · pp. 19–20

7.4 Future Directions in Abstractive Summarization

• Future work should evaluate factual hallucinations, address the harder problem of extrinsic hallucinations, and study dialogue summarization, where 74% of QMSum samples contain inconsistent facts.The survey notes that factual hallucinations may improve summaries despite adding source-absent information, while dialogue requires handling complex discourse relations and speaker coreference.Source: §7.4 · p. 20

8.1 Hallucination Definition in Dialogue Generation

• Intrinsic hallucination contradicts dialogue history or external knowledge, whereas extrinsic hallucination produces responses that cannot be verified from the available inputs.Dialogue systems may tolerate limited open-domain invention, but factual contradictions and unverifiable claims remain problematic.Source: §8.1 · pp. 20–21

8.2 Open-domain Dialogue Generation

• Open-domain systems face self-inconsistency across utterances and external factual inconsistency, with persona consistency and knowledge grounding central to reliable dialogue.Knowledge-grounded dialogue distinguishes intrinsic errors from misused knowledge and extrinsic errors absent from the provided knowledge; implicit language-model knowledge leaves extrinsic hallucination poorly investigated [60].Source: §8.2 · p. 21, §8.2.1 · p. 21, §8.2.2 · p. 22

8.2.1 Self-Consistency.

• PersonaChat conditioning, NLI, and reinforcement learning improve persona consistency, but methods independent of predefined persona descriptions remain necessary [47].Persona descriptions are not always available and cannot cover every aspect of a persona.Source: §8.2.2 · p. 21

8.2.3 Hallucination Metrics.

• No standard metric exists for dialogue hallucination, so evaluation commonly combines human factual-consistency judgments with overlap, NLI-based, trainable, and fact-checking metrics [223].KF1 measures overlap with gold knowledge and RF1 emphasizes infrequent words, while BEGIN and Conv-FEVER support NLI-style consistency evaluation,,, [94], [226].Source: §8.2.3 · p. 22

8.2.4 Mitigation Methods.

• Retrieval, knowledge-aware control codes, inductive attention, response refinement, and RHO’s grounding and reranking mechanisms are used to reduce hallucination in open-domain dialogue,,, [60],.Retrieval substantially improves knowledge-grounded performance and reduces hallucination without sacrificing conversational ability.Source: §8.2.4 · p. 23

8.3 Task-oriented Dialogue Generation

• Task-oriented hallucination can arise when dialogue acts are verbalized incorrectly, including missing or redundant slots, motivating slot error rate and tree accuracy metrics [10].Tree-structured representations with constrained decoding and reinforcement learning with slot-consistency rewards provide corresponding mitigation strategies [10].Source: §8.3 · p. 23, §8.3.1 · p. 23, §8.3.2 · p. 23

8.3.2 Mitigation Methods.

• End-to-end task-oriented systems may generate unverifiable entities absent from the knowledge base, but extrinsic hallucination in these systems remains less explored.A response mentioning an unsupported hotel illustrates this risk.Source: §8.3.2 · pp. 23–24

8.4 Future Directions in Dialogue Generation

• Future work should address self-contradiction caused by truncated or short dialogue memory and develop fact-checking for verifiable-claim detection and evidence retrieval [223], [94].Fact-checking could support both factual-consistency evaluation and dialogue-system modeling.Source: §8.4 · p. 24

9 HALLUCINATION IN GENERATIVE QUESTION ANSWERING

• GQA generates abstractive answers from multiple retrieved documents, making hallucination common when sources are redundant, complementary, contradictory, or irrelevant.Hallucinated answers can mislead users and substantially damage system performance.Source: §9 · p. 24

9.1 Hallucination Definition in GQA

• GQA lacks a standard hallucination definition, but human-rated factual faithfulness, semantic drift, and intrinsic or extrinsic errors provide prevailing conceptualizations.Intrinsic hallucination conflicts with source content, whereas extrinsic hallucination cannot be verified from the source documents.Source: §9.1 · pp. 24–25

9.2 Hallucination-related Metrics in GQA

• No automatic metric specifically evaluates GQA hallucination: ROUGE and F1 correlate poorly with human judgments, while semantic overlap, factual accuracy, factual consistency, and answer recall offer partial alternatives.Human faithfulness judgments remain common but usually cover only small data samples; Su et al. [237] report that answer-recall correlation with human evaluation remains unverified.Source: §9.2 · p. 25

9.3 Hallucination Mitigation in GQA

• GQA hallucination arises from both retrieval failures and conditional-generation errors, motivating knowledge integration, rationale extraction, sparse attention, and machine-reading-augmented generation.These approaches improve use of retrieved information but often require high-quality resources or additional extraction models; simply scaling models is less promising than fine-tuning for truthfulness [155].Source: §9.3 · p. 26

9.4 Future Directions in GQA

• Future GQA research should develop holistic hallucination metrics, hallucination-annotated datasets, stronger query-relevant retrieval, and generators that synthesize accurate answers from multiple sources.Current GQA datasets lack hallucination annotations, and existing metrics capture only individual aspects such as semantic overlap, faithfulness, or factual consistency.Source: §9.4 · pp. 26–27

10.1 Hallucination Definition in Data-to-Text Generation

• Data-to-text hallucinations arise when generated descriptions contradict structured inputs or add irrelevant information, making errors particularly consequential in real-world applications.Intrinsic hallucinations contradict input data, whereas extrinsic hallucinations introduce information absent from the input [45].Source: §10.1 · p. 27, §10 · p. 27

10.2 Hallucination Metrics in Data-to-Text Generation

• Hallucination metrics evaluate faithfulness through table-text alignment, entity coverage, question answering, textual entailment, or language-model loss.These include PARENT and PARENT-T [45], entity-centric coverage and hallucinated ratios, Data-QuestEval, NLI checks [57], and LM-based token ratios [251].Source: §10.2 · pp. 27–28

10.3 Hallucination Mitigation in Data-to-Text Generation

• Mitigation methods improve grounding through faithful datasets, data cleaning, explicit planning or skeletons, faithfulness objectives, controllable hallucination, and uncertainty-aware decoding.UABS penalizes predictive uncertainty during beam search, while Multi-Branch Decoder controls content, hallucination, and fluency using word-level alignments; methods can be combined.Source: §10.3 · pp. 28–29

10.4 Future Directions in Data-to-Text Generation

• Future work should explicitly model numbers because numerical errors frequently cause hallucinations, while current models and metrics rarely treat numbers specially.Better numerical presentation and scale understanding could reduce errors caused by misunderstanding tabular or textual numbers.Source: §10.4 · p. 29
• Reasoning over logical data-to-text inputs, including inference, calculation, comparison, and graph-based numerical reasoning, may improve entity relationships and reduce hallucinations.Logical data-to-text generation is more challenging than surface-level generation because it requires explicit reasoning.Source: §10.4 · p. 29

11 HALLUCINATIONS IN NEURAL MACHINE TRANSLATION

• NMT outputs are typically fluent and human-level, making hallucinations especially dangerous because they can misinform users.Source: §11 · p. 29

11.1 Hallucinations Definition and Categories in NMT

• Early comparisons found that NMT may sacrifice adequacy for fluency, particularly on out-of-domain test sets, and later work treats source-disconnected translations as hallucinations.Source: §11.1 · pp. 29–30
• Intrinsic hallucinations contradict source information, such as translating “Mike” as “Jerry” or omitting the source negation.Source: §11.1 · p. 30
• Extrinsic hallucinations add content without apparent connection to the input, such as “happily” or “with his friend.”Source: §11.1 · p. 30
• Raunak et al. [215] distinguish perturbation hallucinations from natural hallucinations, with the latter divided into detached and oscillatory forms.Perturbation hallucinations arise when perturbed and unperturbed inputs produce drastically different content; detached outputs are semantically disconnected, while oscillatory outputs repeat n-grams.Source: §11.1 · p. 30
• Over-translation repeats words absent from the source, whereas under-translation skips words that require translation.Tu et al. and Kong et al. analyze repetitive output as over-translation.Source: §11.1 · p. 30
• Additional hallucinatory behaviors include abrupt jumps to sequence endings and outputs that remain mostly in the source language.Source: §11.1 · p. 30

11.2 Hallucination Metrics in NMT

• Hallucinations can leave BLEU unchanged, motivating statistical methods that automate their identification and quantification [251].Researchers have traditionally identified hallucinations manually because their definition is qualitative and subjective.Source: §11.2 · pp. 30–31
• BVSS detects adequacy loss when the reference contains more information than the translation, or vice versa.The metric is based on bag-of-vectors sentence similarity.Source: §11.2.1 · p. 31

11.2.2 Model-Based Metrics.

• An auxiliary evaluation decoder estimates faithfulness from word-level translation probabilities alongside fluency, using its loss to adjust translation probabilities.Faithfulness is treated interchangeably with adequacy.Source: §11.2.2 · p. 31
• Average-attention entropy distinguishes hallucinating from non-hallucinating models, because hallucinations concentrate attention on one input point rather than across the sequence [257, 80, 133].The method compares clean and deliberately perturbed test sets when ground-truth translations are unavailable.Source: §11.2.2 · p. 31
• Token-level detectors label synthetic hallucinated tokens and compare prediction loss, while similarity and overlap methods flag tokens lacking source alignment or coverage.CDR is especially effective for under-translation, whereas overlap heuristics can fail with paraphrases and synonyms; embeddings support more diverse synonymous translations.Source: §11.2.2 · pp. 31–32
• Approximate Natural Hallucination detection uses low unique-bigram counts and repeated target sequences to identify oscillatory hallucinations [215].It assumes hallucinations can arise when distinct sources align to the same target.Source: §11.2.2 · p. 32

11.3 Hallucination Mitigation Methods in NMT

• NMT hallucinations motivate mitigation because they are difficult for non-fluent target-language readers to detect and can cause serious errors or dangers.Source: §11.3 · p. 32

11.3.1 Data-Related.

• Data augmentation, especially prefixing common-token insertions, and corpus filtering of repetitive or mismatched pairs are prominent mitigation strategies, though augmentation requires identifying hallucination types.Cross-entropy filtering penalizes bilingual pairs where translation models disagree, while semantic-factor tags identify divergent tokens and can improve overall NMT performance [19].Source: §11.3.1 · p. 32
• Overexposure bias arises because teacher-forced models train on ground truth but infer from potentially incorrect past predictions.Proposed remedies include minimum risk training, differentiable scheduled sampling, and alignment-aware scheduled sampling for differing source-target word orders.Source: §11.3.2 · p. 32

11.3.2 Modeling and Inference.

• Hallucination-aware self-training discards hallucinated-token losses and masks their decoder hidden states, improving BLEU and substantially reducing hallucinations, particularly for low-resource languages.The approach uses token loss truncation and decoder HS masking based on hallucination labels.Source: §11.3.2 · p. 33
• Other modeling and inference remedies include tilted empirical risk minimization, regularization and clipping, and phrase-alignment improvements that support translation accuracy and source-content identification.Source: §11.3.2 · p. 33

11.4 Future Directions in NMT

• Future NMT research should define a metric-based cutoff between translation error and hallucinated content, while repurposing constrained decoding to mitigate hallucinations.A fluency-adequacy threshold is a partial precedent but does not isolate hallucinations, and terminology-constrained decoding methods may be adapted for NMT [10] [49] [236].Source: §11.4 · p. 33
• Future work should simplify computationally expensive hallucination detection and faithfulness-evaluation methods, remove extra perturbed test-set or multi-decoder requirements, and test methods across CNN and Transformer architectures.Existing approaches include Feng et al., Lee et al., and Raunak et al. [215].Source: §11.4 · p. 33

12 HALLUCINATION IN VISION-LANGUAGE GENERATION

• Vision-language models align visual and linguistic representations in a shared multimodal space, so they can hallucinate for reasons similar to NLG systems [51].Models are commonly pretrained either from scratch on image-text pairs or by adapting a pretrained language model to vision-language data.Source: §12 · p. 33
• Research on measuring and mitigating vision-language hallucination remains in its early stage and is largely established only for image captioning.The section presents image-captioning research first, then hallucination phenomena in other tasks and future directions.Source: §12 · pp. 33–34

12.1 Object Hallucination in Image Captioning

• Object hallucination in image captioning comprises intrinsic errors involving nonexistent or incorrect objects and extrinsic errors involving objects whose presence cannot be verified from the image.Rohrbach et al. introduce CHAIR, with CHAIR_i measuring hallucinated object instances and CHAIR_s measuring captions containing at least one hallucinated object.Source: §12.1 · p. 34
• Proposed mitigation methods reduce image-captioning object hallucination through more uniform object co-occurrence statistics, uncertainty-aware beam search, and object masked language modeling [17] [42].These approaches target data augmentation, decoding uncertainty, and vision-language pretraining objectives, respectively; the augmentation method reduces hallucination without changing model architecture [17].Source: §12.1 · pp. 34–35

12.2 Hallucination in Other VL Tasks

• In open-ended visual question answering, models can generate plausible text-only answers that conflict with the image or invent unsupported answers when adversarially prompted with unanswerable questions.These behaviors demonstrate hallucination beyond image captioning and remain an open research question.Source: §12.2 · p. 35

12.3 Future Directions in VL

• Future VL research needs broader empirical analyses, more general hallucination metrics, and controlled generation with visual grounding [42].CHAIR requires a predefined object-category list and does not generalize well, while no automatic metrics currently quantify the hallucination types discussed for other VL tasks.Source: §12.3 · p. 35

13.1 Hallucination Definition in LLMs

• LLM hallucination includes deviations from both input sources and training data, so it is more extrinsic than hallucination in earlier NLG models.Because training data may contain non-factual or contradictory information, hallucination should be distinguished from factuality.Source: §13.1 · p. 36

13.2 Hallucination Metrics for LLMs

• LLM hallucination metrics comprise reference-dependent benchmarks and reference-free detectors based on uncertainty or consistency.Reference-dependent methods use verified knowledge sources, whereas reference-free methods rely on token probabilities or self-consistency.Source: §13.2 · p. 36, §13.2.1 · p. 36, §13.2.2 · p. 37

13.2.1 Reference-dependent Metrics.

• Human-annotated benchmarks target long-tail, adversarial, domain-specific, recent, and uncertainty-aware knowledge, while automated methods generate QA data or retrieve evidence for verification.Automated pipelines can add stronger verification models, tool interaction, claim decomposition, fine-grained retrieval, or chain-of-verification.Source: §13.2.1 · p. 37

13.2.2 Reference-free Metrics.

• Reference-free detection is useful without ground truth: uncertainty methods aggregate token probabilities or entropy, whereas consistency methods compare multiple sampled completions and are more reliable in black-box settings.Consistency can use variation ratios, BERTScore, n-gram or named-entity overlap, NLI, or LLM judgment.Source: §13.2.2 · pp. 37–38

13.3.1 Data-Related Methods.

• Mitigation begins with data quality: credible, filtered, deduplicated, and carefully curated pretraining, instruction-tuning, and reward-model data can reduce hallucination.GPT-4 training used user-flagged non-factual data for open-domain hallucinations and synthetic data for closed-domain reward-model training.Source: §13.3.1 · pp. 38–39

13.3.2 Modelling and Inference Methods.

• Modelling and inference methods include RLHF, model editing, contrastive or intervention-based decoding, reasoning and verification, post-processing, and ensemble debate.RLHF reduced InstructGPT’s closed-domain hallucination rate from 41% to 21%, but alignment can cause capability loss and catastrophic forgetting.Source: §13.3.2 · pp. 39–40

13.4.1 Hallucination in Large Multimodal Models.

• Multimodal hallucination remains difficult because LMMs must represent objects, attributes, spatial relations, and higher-level responses despite imperfect visual representations, alignment, supervision, and costly annotation.Large-scale self-supervision is suggested because supervised learning has been insufficient for robust visual and language systems.Source: §13.4.1 · pp. 40–41

13.4 Future Directions of Hallucination Mitigation in LLMs

• Future work seeks generalizable mitigation for long-tail and low-resource settings, calibrated knowledge-boundary uncertainty, lower alignment tax, and theoretical explanations of hallucination mechanisms.Current challenges include admitting “I don’t know,” expressing graded uncertainty, preserving creativity and abilities, and explaining scaling, generation, SFT, and RLHF effects.Source: §13.4.2 · p. 41, §13.4.3 · p. 41, §13.4.4 · p. 41

14 CONCLUSION

• The survey identifies intrinsic and extrinsic hallucinations as distinct categories requiring different mitigation strategies.It attributes hallucinations to noisy data, erroneous parametric knowledge, incorrect attention, inappropriate training, and inference exposure bias.Source: §14 · p. 42
• Hallucination is concerning because fluent but misleading outputs can harm users in some scenarios and tasks.Source: §14 · p. 42
• Hallucinations are relatively easy to detect in abstractive summarization and neural machine translation when outputs can be checked against source evidence.Source: §14 · p. 42
• Task-specific challenges differ: dialogue requires balancing diversity and consistency, data-to-text reflects input–output format discrepancies, and GQA/VL mitigation remains preliminary.Hallucination harms GQA and VL performance, while visual-language research has mainly focused on object hallucination and unfaithful outputs relative to visual scenes.Source: §14 · p. 42
• Most NMT mitigation methods reduce dataset noise or alleviate exposure bias, while substantial challenges remain in identifying and mitigating hallucinations across NLG.Source: §14 · p. 42