PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions

Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, Ting Liu

arXiv:2311.05232v2cs.CL

TL;DR

LLMs can generate plausible yet factually unsupported content, creating reliability concerns for information retrieval systems. This survey organizes LLM hallucinations through a tailored taxonomy and reviews their causes, detection, benchmarks, mitigation strategies, and open directions, concluding that hallucination remains an ongoing concern requiring continued investigation.

  • Problem

    LLM hallucinations generate seemingly plausible yet factually unsupported content, creating reliability concerns and motivating a comprehensive understanding of their causes and research advances.

    Source: §1 · p. 2

  • Method

    The survey introduces a taxonomy distinguishing factuality and faithfulness hallucinations and synthesizes causes, detection methods, benchmarks, mitigation strategies, retrieval-augmented LLM limitations, and future directions.

    Source: §1 · p. 2, §4.1 · p. 12, §8 · p. 34

  • Results

    The survey’s synthesis identifies hallucination in LLMs as an ongoing concern despite significant research progress in causes, detection, and mitigation.

    Source: §8 · p. 34

  • Takeaways & Limitations

    The survey provides insights intended to guide research toward more robust information retrieval systems and trustworthy artificial intelligence.

    Source: §8 · p. 34

    Current research gives limited attention to capability misalignment, leaving knowledge boundaries in capability alignment insufficiently understood for effective hallucination mitigation.

    Source: §5.2.2 · p. 23

Abstract

from arXiv · show

The emergence of large language models (LLMs) has marked a significant breakthrough in natural language processing (NLP), fueling a paradigm shift in information acquisition. Nevertheless, LLMs are prone to hallucination, generating plausible yet nonfactual content. This phenomenon raises significant concerns over the reliability of LLMs in real-world information retrieval (IR) systems and has attracted intensive research to detect and mitigate such hallucinations. Given the open-ended general-purpose attributes inherent to LLMs, LLM hallucinations present distinct challenges that diverge from prior task-specific models. This divergence highlights the urgency for a nuanced understanding and comprehensive overview of recent advances in LLM hallucinations. In this survey, we begin with an innovative taxonomy of hallucination in the era of LLM and then delve into the factors contributing to hallucinations. Subsequently, we present a thorough overview of hallucination detection methods and benchmarks. Our discussion then transfers to representative methodologies for mitigating LLM hallucinations. Additionally, we delve into the current limitations faced by retrieval-augmented LLMs in combating hallucinations, offering insights for developing more robust IR systems. Finally, we highlight the promising research directions on LLM hallucinations, including hallucination in large vision-language models and understanding of knowledge boundaries in LLM hallucinations.

1 INTRODUCTION

This survey examines LLM hallucinations as plausible but unsupported content that threatens reliable real-world information retrieval, proposing a tailored taxonomy and comprehensive review of causes, detection, mitigation, RAG limitations, and future directions.

Motivation

  • LLMs advance language understanding, generation, reasoning, and information seeking, but convincing hallucinations complicate deployment in decision-relevant systems such as chatbots, search engines, and recommender systems.Hallucinations can spread false beliefs or cause harm when system-provided information influences decisions.Source: §1 · p. 2

Taxonomy

  • The survey defines factuality hallucination as divergence from verifiable facts and faithfulness hallucination as inconsistency with instructions, context, or the generated content itself.Faithfulness hallucination comprises instruction inconsistency, context inconsistency, and logical inconsistency.Source: §1 · p. 2

Survey scope

  • It organizes hallucination causes across data, training, and inference stages, while surveying detection methods, benchmarks, and mitigation approaches aligned with those causes.The benchmarks serve as testbeds for measuring hallucinations and detection effectiveness.Source: §1 · pp. 2–3

RAG and future directions

  • The survey analyzes why retrieval-augmented generation remains vulnerable to hallucinations and identifies future directions including vision-language hallucinations and LLM knowledge boundaries.These analyses aim to inform development of more robust retrieval-augmented systems.Source: §1 · p. 3

Comparing with Existing Surveys

  • Compared with prior surveys, this work uniquely combines a layered hallucination taxonomy, comprehensive causal analysis, and mitigation strategies directly connected to identified causes.It extends factuality-focused trustworthiness discussions to faithfulness hallucinations and distinguishes itself from surveys centered on NLG, mitigation, trustworthiness, or factuality.Source: §1 · p. 3

Organization of this Survey

  • The survey proceeds from taxonomy and causes to detection and benchmarks, mitigation, retrieval-augmented generation challenges, and future research directions.These topics correspond to Sections 2 through 7 and are summarized in Fig 1.Source: §1 · p. 3

2 DEFINITIONS

This section defines LLMs, outlines their three-stage training process, and introduces hallucination as generated content that is nonsensical, unfaithful, contradictory, or unverifiable. It further proposes a taxonomy distinguishing factuality hallucinations from faithfulness hallucinations, each with specific subtypes.

2.1 Large Language Models

  • LLMs are general-purpose transformer-based models trained on massive text corpora, whose scaling enables capabilities such as in-context learning, chain-of-thought prompting, and instruction following.Examples include GPT-3 [29], PaLM [54], LLaMA [300], GPT-4 [232], and Gemini [259].Source: §2.1 · p. 3

2.2 Training Stages of Large Language Models

  • LLMs undergo pre-training, supervised fine-tuning, and reinforcement learning from human feedback, with each stage contributing distinct capabilities relevant to hallucination origins.Pre-training acquires language, world knowledge, and reasoning through next-token prediction; SFT uses annotated instruction-response pairs; RLHF optimizes outputs using human-preference rewards, often with PPO [270].Source: §2.2 · p. 4, §2.2.1 · p. 4, §2.2.2 · p. 5

2.2.1 Pre-training.

  • Pre-training primarily optimizes completion, so pre-trained LLMs can remain misaligned with users’ desired responses until instruction-focused fine-tuning improves controllability and generalization.Pre-training uses autoregressive prediction over extensive corpora, while SFT trains on meticulously annotated instruction-response pairs [370].Source: §2.2.1 · p. 4, §2.2.2 · p. 5

2.2.2 Supervised Fine-Tuning.

  • RLHF aligns LLM outputs with human preferences by training a preference model on ranked responses and maximizing its reward through reinforcement learning [55] [233].This process guides models toward high-quality and harmless responses, typically using Proximal Policy Optimization (PPO) [270].Source: §2.2.3 · p. 5

2.3 Hallucinations in Large Language Models

  • In NLP, hallucination denotes generated content that appears nonsensical or unfaithful to source content, traditionally divided into intrinsic and extrinsic hallucinations [89] [126] [136] [174].Intrinsic outputs conflict with provided context, whereas extrinsic outputs cannot be verified using the context or external knowledge bases.Source: §2.3 · p. 5
  • Because LLM hallucinations arise in open-domain, user-centric interactions and predominantly concern factuality, the survey introduces a refined taxonomy of factuality and faithfulness hallucinations based on Ji et al. [136].Table 1 provides examples for each category.Source: §2.3 · p. 5, §2.3.1 · p. 5, Table 1 · p. 6

2.3.1 Factuality Hallucination.

  • Factuality hallucinations comprise factual contradiction and factual fabrication, including entity-error, relation-error, unverifiability, and overclaim hallucinations.Contradictions contain facts that conflict with real-world information, while fabrications contain claims unverifiable against established knowledge.Source: §2.3.1 · pp. 6–7

2.3.2 Faithfulness Hallucination.

  • Faithfulness hallucinations include instruction inconsistency, context inconsistency, and logical inconsistency, covering deviations from directives, conflicts with provided context, and contradictions within reasoning or answers.These errors can involve performing the wrong task, contradicting supplied information, or producing a final answer inconsistent with the reasoning chain.Source: §2.3.2 · p. 7

3 HALLUCINATION CAUSES

LLM hallucinations arise from multifaceted causes spanning the full capability-acquisition process and are organized into data, training, and inference factors.

3 HALLUCINATION CAUSES

  • The section frames hallucinations as having multifaceted origins across the entire LLM capability-acquisition process.These causes are examined as root causes of hallucinations in LLMs.Source: §3 · p. 7

3.1 Data

  • Data is identified as one of the three primary aspects contributing to LLM hallucinations.Data-related causes are discussed in §3.1.Source: §3 · p. 7

3.2 Training

  • Training is identified as one of the three primary aspects contributing to LLM hallucinations.Training-related causes are discussed in §3.2.Source: §3 · p. 7

3.3 Inference

  • Inference is identified as one of the three primary aspects contributing to LLM hallucinations.Inference-related causes are discussed in §3.3.Source: §3 · p. 7

3.1 Hallucination from Data

Training data expand LLM capabilities while also driving hallucinations through flawed pre-training sources, bounded knowledge, and inferior alignment data. The section highlights misinformation and bias amplification, gaps in long-tail, current, and copyright-restricted knowledge, and instruction-tuning effects.

3.1 Hallucination from Data

  • Pre-training and alignment data build LLM capabilities but also principally contribute to hallucinations through misinformation and bias, knowledge boundaries, and inferior alignment data.Pre-training data provide general capabilities and factual knowledge, whereas alignment data teach instruction following and human-preference alignment [388] [322].Source: §3.1 · p. 7

3.1.1 Misinformation and biases.

  • Flawed pre-training data can cause imitative falsehoods and propagate societal biases, including gender- and nationality-linked stereotypes that create context-inconsistent outputs.Heuristic collection of large web corpora introduces misinformation [328], which memorization can reproduce; biased internet texts can similarly be propagated.Source: §3.1.1 · p. 8

3.1.2 Knowledge Boundary.

  • LLMs hallucinate beyond knowledge boundaries because they miss long-tail facts and lack rapidly changing or copyright-restricted information.Domain-specific medical [179] [279] and legal [353] questions expose long-tail deficits, while inaccessible copyrighted materials create additional knowledge gaps [215].Source: §3.1.2 · pp. 9–10
  • Supervised fine-tuning struggles to add new factual knowledge, and acquiring such knowledge correlates with increased hallucinations; task-format, complex, and diverse instructions further increase hallucinations.These findings concern SFT instruction pairs and instruction effects reported by Gekhman et al. [98] and Li et al. [168].Source: §3.1.3 · p. 10

3.2 Hallucination from Training

Hallucinations can arise from shortcomings across LLM training stages, including pre-training’s causal generation and exposure bias, SFT’s knowledge-boundary misalignment, and RLHF-driven sycophancy. These failures can cause cascading errors, fabrication beyond model knowledge, or outputs that favor user approval over truthfulness.

3.2 Hallucination from Training

  • Training-stage shortfalls can induce hallucinations because pre-training builds representations and world knowledge while alignment shapes responses to user instructions and preferences.Deficiencies in either stage can undermine the capabilities needed for reliable generation.Source: §3.2 · p. 10

3.2.1 Hallucination from Pre-training.

  • Causal left-to-right pre-training can limit contextual dependency modeling, while soft attention and exposure bias contribute to reasoning hallucinations and cascading errors from erroneous tokens.Exposure bias creates a mismatch between training and inference, allowing one generated error to propagate through later tokens like a snowball effect.Source: §3.2.1 · pp. 10–11
  • SFT can amplify hallucinations when instructions exceed pre-training capability boundaries, over-fitting encourages fabrication, and forced completion prevents accurate uncertainty expression.These models are therefore more likely to fabricate responses rather than reject queries beyond their knowledge boundaries [98] [341] [362].Source: §3.2.2 · p. 11

3.2.2 Hallucination from Supervised Fine-tuning.

  • RLHF-trained models may exhibit sycophancy by prioritizing human approval over truthfulness, including endorsing clearly incorrect answers they internally recognize as inaccurate [327] [274].The tendency is linked to human and preference-model biases toward sycophantic responses, despite internal beliefs that may encode statement truthfulness [13] [31] [233].Source: §3.2.3 · p. 11

3.3 Hallucination from Inference

Hallucinations can arise during inference because decoding shortcomings, insufficient context attention, constrained output distributions, and reasoning failures distort otherwise capable LLM generation. Randomness increases diversity but also raises hallucination risk, while over-confidence and reasoning limitations undermine faithfulness and accuracy.

3.3 Hallucination from Inference

  • Higher decoding randomness and temperature increase hallucination risk by sampling lower-frequency tail tokens, despite avoiding the likelihood trap [5] [78] [84] [118] [209].Stochastic sampling supports creative and diverse generation, but diversity is positively correlated with hallucinations.Source: §3.3.1 · p. 11, §3.3 · p. 11

3.3.2 Over-confidence.

  • Localized attention and insufficient context use promote faithfulness hallucinations, while long responses heighten instruction forgetting and divergence from the source context [19] [46] [193].Over-confidence prioritizes partially generated content and fluency over faithful adherence to the original context [45] [212].Source: §3.3.2 · pp. 11–12
  • The Softmax bottleneck limits output-distribution expressivity, and multimodal desired distributions can prevent correct prioritization across modes, increasing hallucination risk.Softmax-based language models may therefore fail to output the desired context-conditioned distribution.Source: §3.3.3 · p. 12

3.3.3 Softmax Bottleneck.

  • Reasoning limitations can cause errors despite relevant knowledge, including multi-hop failures and the Reversal Curse, where reversing a learned relation defeats logical deduction [22] [386].These failures show that effective knowledge use depends on reasoning capabilities, not knowledge possession alone.Source: §3.3.4 · p. 12

4 HALLUCINATION DETECTION AND BENCHMARKS

The section organizes hallucination detection into factuality and faithfulness assessment, then surveys benchmarks for measuring hallucination prevalence and detector performance. It covers retrieval, uncertainty, behavioral, overlap, classifier, QA, and LLM-based methods alongside benchmarks targeting long-tail knowledge, imitative falsehoods, and realistic detection scenarios.

4.1 Hallucination Detection

  • Hallucination detection divides into factuality detection, which identifies real-world inaccuracies, and faithfulness detection, which evaluates adherence to provided context.The section further distinguishes hallucination evaluation benchmarks from hallucination detection benchmarks.Source: §4.1 · p. 12, §4.2 · pp. 16–17

4.1.1 Factuality Hallucination Detection.

  • Factuality detection uses external fact-checking or internal uncertainty estimation, with fact-checking extracting and verifying claims against trusted sources or model knowledge.Uncertainty methods use internal states such as token probability and entropy, or behavior such as self-consistency, indirect questioning, and multi-agent examination; internal knowledge alone may be unreliable.Source: §4.1.1 · pp. 12–13

4.1.2 Faithfulness Hallucination Detection.

  • Faithfulness metrics comprise fact-overlap, classifier, QA, uncertainty, and LLM-based approaches that compare generated content with its source or evaluate model confidence.Fact metrics progress from n-gram to entity and relation overlap; QA metrics compare source and target answers, while LLM judges provide binary or Likert-scale assessments through prompting.Source: §4.1.2 · pp. 14–16
  • Classifier-based faithfulness detection is limited by granularity mismatches, scarce annotations, and gaps between related-task data and downstream inconsistency detection.Researchers address limited supervision by constructing synthetic training data through rule-based perturbation or generation.Source: §4.1.2 · pp. 15–16

4.2.1 Hallucination Evaluation Benchmarks.

  • Hallucination evaluation benchmarks emphasize long-tail knowledge and challenging questions, using multiple-choice accuracy or generative QA scored by humans or proxy models.Examples cover popularity-based, time-sensitive, domain-specific, and adversarial imitative-falsehood knowledge, including PopQA [204], REALTIMEQA [148], FreshQA [308], TruthfulQA [182], HalluQA [49], and HaluEval 2.0 [168].Source: §4.2.1 · pp. 17–18

4.2.2 Hallucination Detection Benchmarks.

  • Hallucination detection benchmarks span synthetic and human-annotated datasets for sentence-, passage-, and long-document evaluation across general, domain-specific, open-domain QA, and summarization tasks.Prior benchmarks largely target task-specific outputs from less capable models, motivating evaluations that reflect more complex LLM-era scenarios; examples include SelfCheckGPT-Wikibio [213], HaluEval [169], BAMBOO [76], ScreenEval [158], FELM [42], PHD [340], RealHall [90], SAC3 [364], and LSum [85].Source: §4.2.2 · pp. 18–19

5 HALLUCINATION MITIGATION

Section 5 reviews contemporary methods for mitigating LLM hallucinations by categorizing them according to their underlying causes: data-, training-, and inference-related hallucinations.

5 HALLUCINATION MITIGATION

  • The section systematically organizes hallucination-mitigation approaches around three cause-specific categories: data-related, training-related, and inference-related hallucinations.Each category presents tailored solutions for challenges associated with its respective hallucination cause.Source: §5 · p. 19

5.1 Mitigating Data-related Hallucinations

Data-related hallucinations are mitigated through data filtering, model editing, and retrieval-augmented generation, targeting misinformation, biases, and knowledge gaps at different stages. These approaches improve factuality but face scalability, model-impact, retrieval-noise, and knowledge-utilization challenges.

5.1.1 Data Filtering.

  • High-quality data selection, factual-data up-sampling, and deduplication reduce misinformation, bias, and semantic redundancy, but filtering remains costly, difficult to scale, and vulnerable to LLM-generated inaccuracies.Methods include curated academic or specialized data, suffix arrays, hash-based matching, MinHash, and embedding-based SemDeDup [1].Source: §5.1.1 · pp. 19–20

5.1.2 Model Editing.

  • Model editing injects updated knowledge by locating and modifying parameters or predicting weight updates, but large-scale or continuous edits can impair generalization and overall model performance.ROME [210] and MEMIT [211] work well empirically mainly on decoder-only LLMs; MAL-MEN addresses cancellation effects in MEND [218] through least-squares aggregation.Source: §5.1.2 · p. 20

5.1.3 Retrieval-Augmented Generation.

  • Retrieval-augmented generation supplies external knowledge through one-time, iterative, or post-hoc retrieval, mitigating knowledge-gap hallucinations without changing the LLM’s parameters.Iterative retrieval supports multi-step reasoning and long-form generation, while post-hoc methods research, verify, revise, or ensemble answers using retrieved evidence.Source: §5.1.3 · pp. 20–21, Figure 4 · p. 21
  • RAG is general and modular across domains, but irrelevant retrieval can introduce noise, while shallow retriever–generator interaction limits knowledge utilization and motivates adaptive retrieval strategies.Future systems should reduce irrelevant retrieval and dynamically adjust retrieval based on query context and prior interaction performance.Source: §5.1.3 · p. 22

5.2 Mitigating Training-related Hallucination

The survey organizes training-related hallucination mitigation around pretraining and alignment strategies, targeting architectural, objective, and preference-related causes. It emphasizes that pretraining interventions can be computationally intensive, while alignment research must better address capability misalignment and knowledge boundaries.

5.2 Mitigating Training-related Hallucination

  • Training-related hallucinations arise from intrinsic limitations in LLM architectures and training strategies, motivating interventions during pretraining and alignment (SFT & RLHF).Source: §5.2 · p. 22
  • Architectural interventions address unidirectional representations and soft attention through BATGPT’s bidirectional autoregression, encoder-decoder designs, and attention-sharpening regularizers [180] [183].Attention-sharpening regularizers promote sparsity and significantly reduce reasoning hallucinations.Source: §5.2.1 · p. 22

5.2.1 Mitigating Pretraining-related Hallucination.

  • Pretraining methods reduce fragmentation and inconsistency by converting sentences into standalone factual units with TOPICPREFIX and training on related-document sequences [160, 276].In-Context Pretraining alters document order to maximize contextual similarity and encourage reasoning across document boundaries.Source: §5.2.1 · pp. 22–23
  • Pretraining mitigation can yield substantial improvements but often requires computationally intensive architectural or objective changes with limited broad applicability.The survey calls for adaptable, efficient strategies that avoid extensive system overhauls.Source: §5.2.1 · p. 23

5.2.2 Mitigating Misalignment Hallucination.

  • Alignment hallucinations involve capability and belief misalignment, with current work focusing mainly on sycophancy caused by preference judgments that favor approval over truthfulness [274].Mitigations improve preference judgments and models [25], aggregate human preferences [274], fine-tune on opinion-independent synthetic data [327], or subtract activation steering vectors [263, 264].Source: §5.2.2 · p. 23
  • Post-training methods directly mitigate hallucinations, but capability misalignment remains underexplored, requiring research into LLM knowledge boundaries.The survey characterizes post-training as effective while identifying capability alignment as a major research gap.Source: §5.2.2 · p. 23

5.3 Mitigating Inference-related Hallucination

Inference-time mitigation targets hallucination by refining decoding to improve factuality, faithfulness to provided context, and consistency within generated content. The survey organizes these methods into factuality-enhanced and faithfulness-enhanced decoding, while noting trade-offs in diversity, reliability, adaptability, and efficiency.

5.3 Mitigating Inference-related Hallucination

  • Inference-time decoding strategies improve output factuality and faithfulness by refining generation without necessarily requiring computation-intensive retraining.The section examines two advanced strategy families for mitigating inference-related hallucination.Source: §5.3 · pp. 23–24, §5.3.1 · p. 24

5.3.1 Factuality Enhanced Decoding.

  • Factuality-enhanced decoding uses adaptive sampling, activation interventions, layer-wise logit selection, and self-correction to reduce non-factual content.Factuality decoding includes factual-nucleus sampling, ITI, and DoLa, whereas post-editing methods use verification questions or iterative self-reflection; these approaches trade factual accuracy against diversity, informativeness, reliability, and speed.Source: §5.3.1 · p. 24

5.3.2 Faithfulness Enhanced Decoding.

  • Faithfulness-enhanced decoding aligns outputs with provided context through context-aware distributions, dynamic temperature adjustment, hallucination detection, evidence-based revision, and context copying.These context-consistency methods address insufficient attention to source context, but excessive contextual emphasis can reduce diversity and some approaches require additional models that increase computational overhead.Source: §5.3.2 · pp. 25–26
  • Logical-consistency methods improve reasoning faithfulness with distilled rationales, contrastive decoding, inferential bridging, preference objectives, and symbolic chain-of-thought verification.They target inconsistent or shortcut-based intermediate reasoning, especially in multi-step and logical tasks, by connecting context, rationales, answers, and explicit symbolic structure.Source: §5.3.2 · pp. 25–26
  • Faithfulness-enhanced decoding advances contextual alignment and internal consistency, but context-aware strategies may lack adaptive attention mechanisms and impose substantial computational costs.The discussion identifies adaptive context handling and efficiency as central remaining challenges.Source: §5.3.2 · p. 26

6 HALLUCINATIONS IN RETRIEVAL AUGMENTED GENERATION

RAG reduces hallucination risk by supplying external knowledge, but hallucinations persist when retrieval fails or generation does not faithfully use retrieved information. The survey organizes these limitations into retrieval failure and generation bottlenecks, covering queries, sources, retrievers, contextual awareness, and contextual alignment.

6 HALLUCINATIONS IN RETRIEVAL AUGMENTED GENERATION

  • RAG can still hallucinate when outputs are factually inaccurate, misaligned with user queries, or unsupported by retrieved information, primarily because of retrieval failure or generation bottlenecks.External knowledge can reduce hallucinations caused by LLM knowledge boundaries, but retrieval-augmented LLMs remain imperfect.Source: §6 · p. 26

6.1 Retrieval Failure

  • Retrieval failures arise from user-query formulation, unreliable or incomplete sources, and ineffective retrievers, producing downstream hallucinations in the RAG pipeline.The retrieval stage must identify relevant information before generation can use it reliably.Source: §6.1 · p. 26

6.1.1 User Queries.

  • User-query problems include unnecessary retrieval, ambiguity, and complexity, motivating adaptive retrieval, query rewriting, and query decomposition to improve evidence relevance.Adaptive retrieval can use heuristics or model self-awareness; rewriting clarifies ambiguous queries, while decomposition addresses multi-step or multi-aspect questions.Source: §6.1.1 · pp. 27–28

6.1.2 Retrieval Sources.

  • Retrieval sources become unreliable when they contain outdated, incorrect, or LLM-generated content that retrieval systems may favor, motivating quality filtering and credibility-aware generation.Generated content can displace human-authored material and introduce factual errors into retrieved contexts.Source: §6.1.2 · pp. 28–29

6.1.3 Retriever.

  • Retriever performance depends on chunking and embeddings: inappropriate granularity can break semantic integrity, while weak or out-of-domain embeddings can misrepresent query-document relevance.RAPTOR, semantic chunking, proposition-level retrieval, and domain-specific embedding fine-tuning address these limitations.Source: §6.1.3 · p. 29

6.2 Generation Bottleneck

  • Even with relevant retrieval, generation can hallucinate because LLMs inadequately maintain contextual awareness or contextual alignment with the retrieved evidence.Contextual awareness concerns using retrieved information, whereas contextual alignment concerns producing outputs that faithfully reflect it.Source: §6.2 · p. 30, §6.2.1 · p. 30, §6.2.2 · p. 32

6.2.1 Contextual Awareness.

  • Contextual awareness is undermined by noisy or conflicting context and insufficient context utilization, including lost-in-the-middle degradation that weakens attention to evidence in long windows.Mitigations include noise-robust training, context compression, counterfactual training, information-seeking tasks, positional rescaling, and hierarchical summarization.Source: §6.2.1 · pp. 30–31

6.2.2 Contextual Alignment.

  • Contextual alignment requires source attribution and faithful decoding because LLMs may prefer parametric knowledge or generate unfaithfully despite accurate retrieved evidence.Approaches include planning or reflection-based attribution, self-attribution, context-aware decoding, semi-parametric generation, and faithfulness-guided beam backtracking.Source: §6.2.2 · pp. 32–33

7 FUTURE DISCUSSION

The future discussion identifies hallucinations in large vision-language models and knowledge-boundary understanding as key research directions. It emphasizes LVLM reliability challenges, scalable mitigation, and the unresolved reliability of probing LLMs’ internal beliefs.

7.1 Hallucination in Large Vision-Language Models

  • LVLMs exhibit multimodal hallucinations, including image-inconsistent objects, attributes, relationships, contextual omissions, and flawed reasoning, threatening reliability in real-world applications [104] [187] [297].They may also be fooled by strong language priors and inappropriate user inputs, even when visual elements are correctly recognized.Source: §7.1 · p. 33
  • Existing LVLM mitigation uses finetuning, post-hoc correction, expert models, or self-utilization, but often requires annotations, visual experts, extra training, and computation, limiting scalability and generalization.The discussion calls for more universal approaches based on faithful, large-scale visual-text pre-training and alignment.Source: §7.1 · p. 33

7.2 Understanding Knowledge Boundary in LLMs

  • LLMs often lack awareness of their knowledge boundaries, confidently producing falsehoods; proposed probes include multiple-choice correctness probabilities and uncertainty from sentence-set similarity [235] [261] [384].These approaches aim to characterize when models possess or lack answerable knowledge.Source: §7.2 · p. 34
  • Although activation-space structures and recent evidence suggest encoded truthfulness and unanswerability, current lie-detector and belief-probing methods remain unreliable and poorly generalizing [13] [31] [220] [281].Levinstein and Herrmann [163] therefore conclude that effectively probing LLMs’ internal beliefs requires further research.Source: §7.2 · p. 34

8 CONCLUSION

The survey examines the causes, detection methods and benchmarks, and mitigation strategies for hallucinations in large language models, while emphasizing that the problem remains unresolved and requires continued investigation.

8 CONCLUSION

  • The survey comprehensively covers LLM hallucination causes, detection methodologies and benchmarks, and mitigation strategies.It aims to support the development of more robust information retrieval systems and trustworthy, reliable, and safe artificial intelligence.Source: §8 · p. 34

Related papers

Loading 2311.05232v2…