PaperlayerView exact revision on arXiv

Source-linked AI summary

Large Language Models Hallucination: A Comprehensive Survey

Aisha Alansari, Hamzah Luqman

arXiv:2510.06265v3cs.CL

TL;DR

LLM hallucinations produce fluent but inaccurate or unsupported content, undermining reliability in applications requiring factual accuracy. This survey synthesizes their causes across the LLM lifecycle and organizes detection, mitigation, evaluation, and explainability research. It concludes that complementary or hybrid techniques are more promising than any single approach, while current methods remain constrained by issues such as computational cost and inconsistent evaluation standards.

  • Problem

    LLM hallucinations are fluent outputs that can be factually inaccurate or unsupported, threatening the reliability of AI-generated content in sensitive applications.

  • Method

    The survey analyzes hallucination causes across the LLM development lifecycle and synthesizes taxonomies of detection, mitigation, explainability, benchmarks, and metrics.

  • Results

    The survey finds that no single detection or mitigation approach performs well across all circumstances, making combinations of complementary techniques the most promising direction.

  • Takeaways & Limitations

    Hybrid approaches combining prompting or reasoning with retrieval and model-centric training or adaptation are identified as promising for improving hallucination robustness and accuracy.

  • Takeaways & Limitations

    Explainability methods can be computationally expensive, and standardized benchmarks and metrics for evaluating explanation quality are lacking.

Abstract

from arXiv · show

Large language models (LLMs) have transformed natural language processing, achieving remarkable performance across diverse tasks. However, their impressive fluency often comes at the cost of producing false or fabricated information, a phenomenon known as hallucination. Hallucination refers to the generation of content by an LLM that is fluent and syntactically correct but factually inaccurate or unsupported by external evidence. Hallucinations undermine the reliability and trustworthiness of LLMs, especially in domains requiring factual accuracy. This survey provides a comprehensive review of research on hallucination in LLMs, with a focus on causes, detection, and mitigation. We first present a taxonomy of hallucination types and analyze their root causes across the entire LLM development lifecycle, from data collection and architecture design to inference. We further examine how hallucinations emerge in key natural language generation tasks. Building on this foundation, we introduce a structured taxonomy of detection approaches and another taxonomy of mitigation strategies. We also analyze the strengths and limitations of current detection and mitigation approaches and review existing evaluation benchmarks and metrics used to quantify LLMs hallucinations. Finally, we outline key open challenges and promising directions for future research, providing a foundation for the development of more truthful and trustworthy LLMs.

1 INTRODUCTION

This survey examines LLM hallucinations as fluent but inaccurate or unsupported content, tracing their causes across development and organizing detection and mitigation research into structured taxonomies. It also reviews benchmarks, metrics, reasoning-aware methods, and multilingual or low-resource challenges.

  • Hallucination and its causes: Hallucinations are fluent, syntactically correct outputs that are factually inaccurate or unsupported by external evidence.They threaten reliability in sensitive applications including medical summarization, financial analysis, and legal counsel.
  • Hallucination and its causes: The survey analyzes hallucination causes across six LLM development stages, from data collection and preparation through inference.The stages also include model architecture, pre-training, fine-tuning, and evaluation.
  • Detection: The survey classifies detection methods into retrieval-, uncertainty-, embedding-, learning-, and self-consistency-based approaches.It discusses the potential of each category for hallucination detection.
  • Mitigation: The survey classifies mitigation methods into prompt-, retrieval-, reasoning-, and model-centric training and adaptation-based approaches.It also reviews chain-of-thought, iterative refinement, and chain-of-verification methods for complex tasks.
  • Evaluation and scope: The survey reviews benchmark datasets and evaluation metrics while identifying limitations in existing evaluation resources.It additionally addresses explainability and challenges specific to multilingual and low-resource settings.

2 RELATED SURVEYS

The survey situates itself among prior work on hallucination taxonomies, causes, detection, and faithfulness evaluation. It distinguishes its contribution through development-cycle analysis, finer-grained detection and mitigation taxonomies, and multilingual coverage.

  • Prior surveys: Prior surveys connected hallucination causes across data, training, and inference stages with detection and mitigation strategies.Other work classified hallucinations as factuality or faithfulness types and linked strategies to foundational causes.
  • Prior surveys: Earlier research also surveyed detection methods and taxonomies spanning intrinsic, extrinsic, factual, contextual, temporal, ethical, and domain-specific hallucinations.These studies additionally considered data, model, and prompt-related causes and faithfulness metrics.
  • Distinctive contributions: Unlike prior surveys, this survey links a detailed taxonomy of hallucination causes explicitly to the LLM development cycle.It also develops more fine-grained taxonomies for hallucination detection and mitigation.
  • Distinctive contributions: The survey extends coverage to multilingual contexts and language-specific challenges often disregarded in existing literature.This broadens its review beyond the taxonomic and methodological coverage of prior surveys.

3 LLMS HALLUCINATION

LLM hallucination is fluent, coherent output that lacks factual grounding, with modern models making such errors harder to detect. The section distinguishes hallucination types and shows how they arise across open-ended generation tasks and critical domains.

  • Definition of Hallucination: LLM hallucination is fluent and coherent text that is factually inaccurate or unsupported by external evidence.
  • Definition of Hallucination: Modern LLM hallucinations are harder to identify than earlier LM errors because improved fluency and contextual relevance can mask inaccuracies.
  • Definition of Hallucination: Creativity produces deliberate, goal-driven novelty, whereas hallucination typically generates unintended factually incorrect or misleading content.
  • Types of Hallucination: Intrinsic hallucination contradicts facts in a source document, while factuality and faithfulness taxonomies capture real-world divergence and drift from instructions or context.
  • Hallucination in NLG Tasks: Hallucinations occur mainly in open-ended NLG tasks, including translation, summarization, generative QA, dialogue, data-to-text, and paraphrasing.
  • Hallucination in NLG Tasks: Data-to-text models can hallucinate because structured data must be converted into natural language, while paraphrasing can add unsupported information or omit crucial details.
  • Impact of Hallucination in Critical Domains: Medical hallucination detection depends heavily on domain expertise and prompt detail, while financial hallucinations can contribute to misallocation of capital.

4 HALLUCINATION CAUSES

Hallucinations arise from interacting weaknesses across the LLM development lifecycle, from data curation and pre-training to fine-tuning and inference. The survey organizes these causes by development stage and shows how data quality, training objectives, alignment, input ambiguity, sampling, and architecture can undermine factual reliability.

  • Lifecycle overview: The survey examines hallucination causes across six development stages, spanning data collection and preparation, architecture, pre-training, fine-tuning, evaluation, and inference.This lifecycle view is intended to identify contributing factors throughout the pathway from pre-training to generation.
  • Data curation: Scaling pre-training data improves capabilities but creates persistent data-quality challenges, while bias, memorization, and over-represented content can produce divergent outputs.Training data may encode societal imbalances and frequently occurring information that models unintentionally acquire or amplify.
  • Architecture and inference: Unidirectional contextualization and SoftMax limitations can leave models reliant on local patterns or poorly represent multiple plausible meanings, contributing to hallucination.These architectural and inference-related limitations become problematic when inputs are ambiguous, incomplete, or contextually rich.
  • Model pre-training: Pre-training can induce hallucinations through shortcut learning, teacher forcing, and insufficient negative examples that weaken factual discrimination.Shortcut learning relies on superficial patterns, while exposure bias arises because training uses flawless contexts but inference uses previously generated tokens; negative examples help reduce plausible false answers.
  • LLM fine-tuning: Fine-tuning and alignment can increase hallucination risk through task-specific overfitting and capability misalignment that encourages confident answers beyond the model’s knowledge.RLHF may prioritize coherence and confidence over factuality when alignment expectations exceed intrinsic capabilities.
  • Inference: During inference, ambiguous prompts can encourage fabricated answers, while stochastic decoding increases the chance of selecting low-probability tokens that diverge from factual correctness.Greedy decoding can reduce hallucination risk but may produce repetitive responses, whereas top-k and nucleus sampling increase creativity alongside factual risk.

5 HALLUCINATION DETECTION

The survey organizes hallucination detection into five categories and reviews their methods, strengths, limitations, and representative developments. It concludes that no single approach performs well universally, motivating combinations of complementary techniques.

  • The taxonomy groups hallucination detection into retrieval-, uncertainty-, embedding-, learning-, and self-consistency-based techniques.
  • Retrieval-Based Detection: Retrieval-based methods compare outputs with trusted external knowledge to assess factual consistency, but depend strongly on retrieval quality.RAG retrieves relevant documents and can be followed by fact-checking models.
  • Uncertainty-Based Detection: Uncertainty-based methods flag low-confidence outputs using token-, semantic-, or structure-level signals without external knowledge retrieval.Examples include normalized token log probabilities, semantic entropy, and attention-derived structural features.
  • Self-Consistency-Based Detection: Self-consistency methods generate multiple responses and assess agreement without ground-truth references, but their effectiveness depends on prompt diversity and sampling strategies.
  • The survey finds that combining complementary approaches is promising because individual methods face distinct weaknesses across knowledge quality, domains, and languages.Retrieval methods handle factual hallucinations effectively but are sensitive to external knowledge, while learning and embedding approaches can degrade out of domain or in low-resource languages.
  • Learning-Based Detection: Learning-based detectors can generalize across hallucination types, but supervised approaches require extensive labeled or synthetic data and may not generalize beyond their training conditions.

6 HALLUCINATION EXPLAINABILITY

The survey distinguishes model-internal and evidence-based explainability for hallucinations. It highlights unresolved tensions among faithfulness, human interpretability, computational practicality, and standardized evaluation.

  • Hallucination explainability seeks human-understandable reasons for why hallucinations occurred and where they originated.
  • Model-Internal Explainability: Model-internal explainability interprets hallucinations through model outputs, uncertainty, attention behavior, and other internal dynamics.RePPL attributes hallucinations to token-level uncertainty linked to semantic propagation instability and low-confidence generation decisions.
  • Grounding Explainability: Grounding explainability examines alignment with external evidence through source documents, factual references, claim decomposition, entailment checking, and structured verification.HaluCheck decomposes outputs into atomic facts and uses retrieved evidence with NLI models to identify unsupported statements.
  • Explainability Challenges: There is no universally agreed correct explanation for an individual hallucination, so current methods often rely on proxy signals that may not fully capture underlying causes.
  • Explainability Challenges: Explanations face a trade-off between faithful technical detail and accessibility to non-experts, while computational expense limits use on long texts and real-time applications.
  • Explainability Challenges: A lack of standardized benchmarks and metrics makes explanation quality difficult to compare and improvement difficult to measure.

7 HALLUCINATION MITIGATION

The survey organizes hallucination mitigation into prompt-, retrieval-, reasoning-, and model-centric training or adaptation-based approaches. It reviews techniques that ground outputs, improve reasoning, verify claims, retrieve relevant information, or alter model behavior during decoding and training.

  • Overview: Hallucination mitigation modifies LLMs to reduce factually inaccurate, ungrounded, and contextually inconsistent outputs.The survey distinguishes mitigation from detection, which identifies hallucinatory outputs rather than changing the model to produce more accurate responses.
  • 7.1 Prompt-Based Techniques: Prompt-based methods guide factual generation through tags, in-context examples, iterative refinement, and up-to-date information.Context-embedded tags achieved a 98.88% success rate in eliminating fabricated information, while iterative refinement and current few-shot examples improved summary faithfulness and factual accuracy.
  • 7.2 Retrieval-Based Techniques: Retrieval-based methods supplement static parametric knowledge with dynamically accessed external or structured knowledge to improve factual grounding.Knowledge graphs can support fact verification, explanations, neuro-symbolic reasoning, constraint alignment, and faithful graph-to-text generation.
  • 7.3 Reasoning-Based Techniques: Reasoning-based methods use structured chains, self-consistency, refinement, verification, inference, and retrieved evidence to reduce logical and factual errors.CoQ requires each reasoning step to have retrieved support and reduces factual inaccuracies by 31% compared with CoT alone; CoVE’s factor+revise strategy achieves the strongest reported overall factuality score among its variants.

8 BENCHMARK DATASETS

The survey reviews datasets for detecting and mitigating hallucinations across NLG tasks, finding strong concentration in QA, summarization, and English. It also highlights emerging multilingual resources and persistent task-, domain-, and benchmark-dependence.

  • Dataset types: Detection datasets pair inputs and outputs with hallucination annotations, while mitigation datasets emphasize factually correct, contextually grounded pairs and external references.Detection data can also support mitigation through negative examples, contrastive learning, prompt design, or generation-time filtering.
  • Coverage: QA and summarization are the predominant evaluated tasks, while English remains the leading language and German, Arabic, Chinese, and multilingual datasets are emerging.Most proposed datasets support both detection and mitigation experiments.
  • Benchmark dependence: Benchmark results remain highly task- and benchmark-dependent, with no single method consistently outperforming others across datasets or under domain and task shifts.Benchmark diversity and evaluation differences complicate direct comparisons.
  • Empirical measurement: Hallucination rates below a few percentage points have been reported for some state-of-the-art models on grounded summarization, but rates vary across model families and prompting behaviors.These results concern narrow, constrained summarization settings rather than a unified cross-task frequency measure.
  • Empirical measurement: Community leaderboards provide an early common reference for comparing model factual consistency, although task-specific benchmarks do not yet yield a unified cross-task hallucination frequency metric.The Hallucination Leaderboard aggregates grounded summarization performance and reports rates computed by HHEM.

9 HALLUCINATION DETECTION AND MITIGATION METRICS

The survey organizes hallucination metrics into statistical, data-driven, human-based, and mixed approaches. These methods trade simplicity, automated coverage, human alignment, and contextual breadth, with no single metric addressing every limitation.

  • Metric taxonomy: Hallucination metrics comprise statistical, data-driven, human-based, and mixed categories, each offering distinct insights while retaining limitations in faithfulness, factuality, or generalizability.The taxonomy spans automated scores, model-based assessments, human judgments, and combinations of multiple signals.
  • Statistical metrics: Statistical metrics compare generated content with references using token overlap, semantic similarity, or uncertainty measures, but often align poorly with human hallucination judgments.Examples include BLEU, ROUGE, BERTScore, and perplexity.
  • Data-driven metrics: Data-driven metrics detect mismatches using curated data or model outputs, including reference-free consistency checks, supervised NLI classifiers, and LLM-as-a-judge frameworks.SelfCheckGPT, FactCC, and G-eval represent these approaches.
  • Data-driven metrics: GPT-4-based evaluation achieves the best overall alignment with human judgment among the described LLM-based evaluation approaches.G-eval converts user-defined criteria into structured reasoning steps and scores outputs through form filling.
  • Human and mixed metrics: Human evaluation assesses dimensions including faithfulness, factuality, consistency, relevancy, and adequacy, while eye tracking uses gaze and pupil responses as additional signals.Mixed approaches combine automated signals with human annotations or domain expertise for context-aware analysis.

10 OPEN ISSUES AND FUTURE DIRECTIONS

Open issues center on weak generalization, attention’s difficulty separating useful context from noise, and benchmark limitations in coverage, granularity, and evaluation reliability. The survey points toward multilingual, more diverse, and more diagnostically informative evaluation and detection methods.

  • Generalization: Detection methods trained on particular datasets and tasks generalize inadequately, while inconsistent training and prompting hinder universal detection frameworks.Multilingual datasets and benchmarks are identified as enabling broader evaluation across linguistic settings.
  • Attention mechanisms: Softmax bottlenecks can prevent attention mechanisms from distinguishing useful context from noise, causing improper weighting of context and self-generated content.The survey proposes investigating adaptive uncertainty measures that dynamically emphasize key contexts.
  • Benchmark coverage and diversity: Benchmarks concentrated on QA, summarization, translation, news, and Wikipedia introduce task and domain bias and provide limited coverage of open-ended, dialogue, and code generation.Dialogue and code-generation hallucination benchmarks remain scarce.
  • Benchmark design: Binary hallucination labels overlook partial errors and varying severity, reducing diagnostic granularity and limiting benchmarks’ value for model improvement.Subtle factual drifts are not distinguished by coarse-grained labeling.
  • Evaluation reliability: Human-dependent ground truths raise scalability, agreement, and bias concerns, while LLM judges can shift preferences in response to visually appealing formatting.The formatting effect can occur even when underlying content quality is unchanged.

11 CONCLUSION

The survey synthesizes hallucination causes, detection techniques, mitigation strategies, benchmarks, and metrics into structured taxonomies. It concludes that progress remains constrained by generalizability, efficiency, multilingual handling, and interpretability, motivating more reliable and factually grounded LLMs.

  • Conclusion: The survey reviews hallucination causes and proposes taxonomies covering five detection dimensions and four mitigation dimensions.It also reviews benchmarks and performance metrics for hallucination detection and mitigation.
  • Conclusion: Current hallucination methods frequently remain limited in generalizability, computational efficiency, multilingual contexts, and interpretability.The survey identifies robust, more generalizable detection methods applicable across domains and model architectures as a future direction.
  • Future direction: Combining advances in model architecture, training techniques, and external knowledge integration is presented as a route toward more reliable and factually grounded LLM outputs.The conclusion frames hallucination mitigation as important for accurate, reliable, and human-values-aligned applications.
Loading 2510.06265v3…