Introduction

In the ever-evolving realm of large language models (LLMs), a constellation of innovative creations has emerged, such as GPT-3 [1], InstructGPT [2], FLAN [3], PaLM [4], LLaMA [5] and other notable contributors [6,7,8,9]. These models implicitly encode global knowledge within their parameters during the pre-training phase [10,11], offering valuable insights as knowledge repositories for downstream tasks [12,13,14]. Nevertheless, the generalization of knowledge can result in memory distortion, an inherent limitation that may give rise to potential inaccuracies [15]. Moreover, their ability to represent knowledge is constrained by model scale and faces challenges in addressing long-tailed knowledge problems [16,17]. While the privacy and timeliness of data in the real world [18,19] unfortunately exacerbate this problem, leaving models difficult to maintain a comprehensive and up-todate understanding of the facts. These challenges present a serious obstacle to the reliability of LLMs, which we refer to as hallucination. [20]. A prominent example of this drawback is that models typically generate statements that appear reasonable but are either cognitively irrelevant or factually incorrect. In light of this observation, hallucinations remain a critical challenge in medical [21,22], financial [23] and other knowledge-intensive fields due to the exacting accuracy requirements. Particularly, the applications for legal case drafting showcase plausible interpretation as an aggregation of diverse subjective perspectives [24].

Definition of Hallucination.

As depicted in Figure 1, hallucination refers to the generation of texts or responses that exhibit grammatical correctness, fluency, and authenticity, but deviate from the provided source inputs (faithfulness) or do not align with factual accuracy (factualness) [25]. In traditional NLP tasks [26], hallucinations are often synonymous with faithfulness: conflicting information leads to Intrinsic Hallucination, i.e., LMs conflict with the input information when generating a response; Conversely, generating ambiguous supplementary information may lead to Extrinsic Hallucination, i.e., LMs produce personal names, historical events, or technical documents that are challenging to verify. LLMs-oriented hallucinations instead prioritize factualness, focusing on whether the result can be evidenced or negated by reference to external facts in the real world. Uncritical trust in LLMs can give rise to a phenomenon Cognitive Mirage, contributing to misguided decision-making and a cascade of unintended consequences [27].

Present work

To effectively control the risk of hallucinations, we summarize recent progress in hallucination theories and solutions in this paper. We propose to organize relevant work by a comprehensive survey (Figure 2):

• Theoretical insight and mechanism analysis. We provide in-depth theoretical and mechanism analysis from three typical perspectives: data collection, knowledge gap and optimization process, reviewing the recent and relevant theories related to hallucinations ( §2). • Taxonomy of hallucination in LLMs. We conduct a comprehensive review of hallucination in LLMs together with a task axis. We review the task-specific benchmarks with a comprehensive comparison and summary ( §3). • Wide coverage on emerging hallucination detection and correction methods. We propose a comprehensive investigation into the proactive detection ( §4) and mitigation of hallucinations handling language pairs with limited resources or non-English translations [39]. Furthermore, cuttingedge Large Vision-Language Models (LVLMs) exhibit instances of hallucinating common objects within visual instructional datasets and prone to objects that frequently co-occur in the same image [40,41].

Knowledge Gap Knowledge gaps are typically attributed to differences in input format between the pre-training and fine-tuning stages [42]. Even when considering the automatic updating of textual knowledge bases, the output can deviate from the expected corrections [43]. For example, questions often do not align effectively with stored knowledge, and the available information remains unknown until the questions are presented. This knowledge gap poses thorny challenges in balancing memory with retrieved evidence, which is construed as a passive defense mechanism against the misuse of retrieval [44]. To delve into this issue, [45] and [46] propose that disregarding retrieved evidence introduces biased model knowledge, while mis-covering and over-thinking disrupt model behavior. Furthermore, in scenarios where a cache component is utilized to offer historical memory during training [47], the model also experiences inconsistency between the present hidden state and the hidden state stored in the cache.

Optimization Process

The maximum likelihood estimation and teacher-forcing training have the potential to result in a phenomenon known as stochastic parroting [48], wherein the model is prompted to imitate the training data without comprehension [49]. Specifically, exposure bias between the training and testing stages have been demonstrated to lead to hallucinations within LLMs, particularly when generating lengthy responses [50]. Besides, sampling techniques characterized by high uncertainty [51], such as top-p and top-k, exacerbate the issue of hallucination. Furthermore, [27] observes that LLMs tend to produce snowballing hallucinations to maintain coherence with earlier hallucinations, and even when directed with prompts as "Let's think step by step", they still generate ineffectual chains of reasoning [13].

Taxonomy of Hallucination

In this paper, we mainly consider representative hallucinations, which are widely observed in various downstream tasks, i.e. Machine Translation, Question and Answer, Dialog System, Summarization System, Knowledge graph with LLMs, and Visual Question Answer. As shown in Table 1, these hallucinations are identified complex taxonomy in numerous mainstream tasks associated with LLMs. In the following sections, we will introduce representative types of hallucinations to be resolved.

• Machine Translation. Since perturbations (e.g., spellings or capital errors) can induce hallucinations reliably, traditional machine translation models tend to validate instances memorised by the model when subjected to perturbations [87,88]. It is worth noting that hallucinations generated by LLMs are mainly translation off-target, over-generation, or failed translation attempts [39]. While in low-resource language setting, most models exhibit subpar performance due to the lack of annotated data [54]. In contrast, they are vulnerable to increased amount of pre-trained languages in multilingual setting [89]. Subsequently, familial LLMs trained on different scales of monolingual data are proved to be viscous [39], as the source of oscillatory hallucination pathology. • Question and Answer. Imperfect responses suffer from flawed external knowledge, knowledge recall cues and reasoning instruction [42]. For example, LLMs are mostly unable to avoid answering when provided with no relevant information, instead provide incomplete and plausible answers [56]. In additon to external knowledge, memorized information without accurate, reliable and accessible source also contributes to different types of hallucinations [22]. Though scaling laws suggest that perplexity on the training distribution is positively correlated with parameter size, [30] further discovers that scaling up models should increase the rate of imitative falsehoods.

• Dialog System. Some studies view dialogue models as unobtrusive imitators, which simulates the distributional properties of data instead of generating faithful output. For example, Uncooperativeness responses [57] originating from discourse phenomena inclines to output an exact copy of the entire evidence. [58] reports more nuanced hallucinations in KG-grounded dialogue systems as analyzed through human feedback. Similarly, FaithDial [59], BEGIN [60], MixCL [61] all implement experiments on the WoW dataset to conduct a meta-evaluation of the hallucination in knowledge grounded dialogue.

• Summarization System. Automatically generated abstracts based on LLMs may be fluent, but they still typically lack faithfulness to the source document. Compared to the human evaluation of traditional summarization models [26], the summarizations generated by LLMs can be categorized into two major types: intrinsic hallucinations that distort the information present in the document; extrinsic hallucinations that provide additional information that cannot be directly attributed to the document [65]. Note that extrinsic hallucination as a metrics of factually consistent continuation of inputs in LLMs is given more attention in summarisation systems [62,64]. Furthermore, [63] subdivides extrinsic hallucinations into factual and non-factual hallucinations. The former provides additional world knowledge, which may benefit comprehensive understanding.

• Knowledge Graph with LLMs. Despite the promising progress in knowledge-based text geneartion, it encounters intrinsic hallucinations inherent to the process where the generated text not only covers the input information but also incorporates redundant details derived from its internal memorized knowledge [90]. To address this, [66] establish a distinction between correctly generated knowledge and knowledge hallucinations in terms of knowledge creation. Notably, the Virtual Knowledge Extraction [91] underscores the potential generalization capabilities of LLMs in the realms of constructing and inferring from knowledge graphs. [32] further empower LLMs to produce interpretable fact-checks through a neural symbolic approach. Based on their fidelity to the source, hallucinations are defined as subject hallucination, relation hallucination, and object hallucination.

• Cross-modal System. Augmented by the superior language capabilities of LLMs, performance of cross-modal tasks achieves promising progress [92,40]. However, despite replacing the original language encoder with LLMs, Large Visual Language Models (LVLMs) [93] still generate object descriptions that not present in the target image, denoted as object hallucinations [41]. Especially, the various failure cases could be typically found in Visual Question Answering [41,67], Image Captioning [94,95,96],

Report Generation [68] etc.

Hallucination Detection

Conventional hallucination detection mainly depends on task-specific metrics, such as ROUGE and BLEU to evaluate the information overlap between source and target texts in summarization tasks [97], while knowledge F1 to estimate the knowledge-aware ability of response generation [98]. These metrics focus on measuring faithfulness of references and fail to provide an assessment of factualness. Despite some reference-free works are proposed, plugin-based methods [99] suffer from world knowledge limitation. QA-based matching metrics [100] lack knowledge completeness of source information. NLI-based methods [60] are unable to support finer-grained hallucination checking as they are sentencelevel, besides entailment and hallucination problems are not equivalent. As the paradigm shift in hallucination detection arising from the rapid development of LLMs, we present a novel taxonomy in Fig 3 and introduce each category in following sections.

• Inference Classifier. The most straightforward strategy involves adopting classifiers to assess the likelihood of hallucinations. Concretely, given a question 𝒬 and an answer 𝒜, an inferential classifier 𝒞 can be asked to determine whether the answer contains hallucinatory content ℋ via computing 𝑝(ℋ) = ℱ 𝒞 (𝒬, 𝒜). Therefore, [64] employs the state-of-the-art LLMs to do end-to-end text generation of detection results. Some other studies [31] finds that adding chains of thought indiscriminately may intervene in the final judgement, whereas retrieving the knowledge properly resulted in gains. Furthering this concept, the hinted classifer and explainer [64], used to generate intermediate process labels and high-quality natural language explanations, are demonstrated to enhance the final predicted class from a variety of perspectives. Subsequently, [62] suggests adopting a different classifier model to the generated model, contributing to easier judgement of factual consistency. For radiology report generation, binary classifiers [68] can be leveraged to measure the reliability by combining image and text embedding. Unlike previous work that employs complex human-crafted rules to evaluate object hallucinations, GAVIE [67] scores responses towards image content based on both accuracy and relevance criteria, which evaluates the LMMs output in an open-ended manner.

• Uncertainty Metric. It is important to examine the correlation between the hallucination metric and the quality of output from a variety of perspectives. One intuitive approach is to employ the probabilistic output of the model itself, as ASTSN [75] calculates the models' uncertainty about the identified concepts by utilising the logit output values. Similarly, BARTSCORE [70] employs a universal notion that models trained to convert generated text to reference output or source text will score higher when the generated text is superior. It is an unsupervised metric that supports the addition of appropriate prompts to improve the measure design, without human judgement to train. Furthermore, KoK [71] based on the work of [101] evaluates answer uncertainty from three categories, i.e., subjectivity, hedges and text uncertainty. However, SLAG [72] measures consistent factual beliefs in terms of paraphrase, logic, and entailment. In addition to this, KLD [73] combines information theory-based metrics (e.g., entropy and KL-divergence) to capture knowledge uncertainty. Beside expert-stipulated programmatic supervision, POLAR [74] introduces Pareto optimal learning assessed risk score for estimating the confidence level of a response.

• Self-Evaluation. To self-evaluate is challenging since the model might be overconfident about its generated samples being correct. The motivating idea of SelfCheckGPT [77] is to use the ability of the LLMs themselves to sample multiple responses and identify fictitious statements by measuring the consistency of information among responses. [76] further illustrates that both the increase in size and the demonstration of assessment can improve self-assessment. Beyond repetitive multiple direct queries, [78] uses open-ended indirect queries and compares their answers to each other for an agreed-upon score outcome. SelfCk [81] imposes appropriate constraints on the same LLM to generate pairs of sentences triggering self-contradictions, which prompt the detection. In contrast, Polling-based querying [41] reduce the complexity of judgement by randomly sampling query objects. Besides, Self-Checker [79] decomposes complex statements into multiple simple statements, fact-checking them one by one. However, [80] introduces two LLMs to drive the complex fact-checking reasoning process by crosscheck.

• Evidence Retrieval. Evidence retrieval accomplishes factual detection by retrieving supporting evidence related to hallucinations. To this end, Designing a claim-centric pipeline allows for a questionretrieve-summary chain to effectively collect original evidence [84,85]. Consequently, FActScore [83] calculates the percentage of atomic facts supported by the given knowledge source. Towards adapting the tasks that users in interaction with generative models, FacTool [86] proposes to integrate a variety of tools into a task-agnostic and domain-agnostic detection framework, in order to assemble evidence about the authenticity of the generated content.

Hallucination Correction

In this section, we delve into the methods to correct hallucination in terms of different aspects. As shown in Figure 4, these hallucination correction paradigms have demonstrated strong dominance in many mainstream NLP tasks. Note that these methods are not entirely orthogonal but could complement each other as required by the tasks in practical applications. In the following sections, we will introduce each methods as shown in Figure 5.

Hallucination Correction

Parameter Adaptation Factual-Nucleus [51], CLR [61], Edit-TA [102], EWR [103], PURR [104], mmT5 [55], HISTALIGN [47], TYE [105], ALLM [106], TRAC [107], Inference-Time [108], EasyEdit [109], DoLa [110] Post-hoc Attribution and Edit Technology NP-Hunter [111], CoT [14], ORCA [112], RR [113], TRAK [114], Data-Portraits [115], Self-Refine [116], Reflexion [117], QUIP [118], Verify-and-Edit [119], CoVe [120], CoNLI [69] Leverage External Knowledge RETRO [121], IRCoT [122], POPQA [17], LLM-AUGMENTER [123], In-Context RALM [62], GeneGPT [124], cTBL [125], CoK [126], FLARE [127], Gorilla [128], RETA-LLM [129], KnowledGPT [130] Assessment Feedback LSHF [131], TLM [132], BRIO [133], LM-know [76], Chain-of-Hindsight [134], ZEROFEC [43], CRITIC [135], VIVID [96], LMH-Snowball [27], MixAlign [45], REFEED [15], PaD [136], ALCE [137], Do-LLM-Know [78], CRL [138], SR [139] Mindset Society HLMTM [39], Multiagent-Debate [140], MAD [141], FORD [142], LM-vs-LM [80], PRD [143], SPP [144] Figure 5: Taxonomy of Hallucination Correction.

• Parameter Adaptation. Parameters in LLMs store biases learned in pre-training, are often unaligned with user intent. A cutting-edge strategy is to guide effective knowledge through parameter conditioning, editing, and optimisation. For example, CLR [61] optimises to reduce the generation probability of negative samples at span level utilising contrastive learning parameters. While introducing contextual knowledge background that contradicts the model's intrinsic prior knowledge, TYE [105] effectively reduces the weight of prior knowledge through context-aware decoding method. Besides, PURR [104] corrupts noise into the text, fine-tune compact editors, and denoise by merging relevant evidence. To introduce additional cache component, HISTALIGN [47] discovers that its hidden state is not aligned with the current hidden state, and proposes sequence information contrastive learning to improve the reliability of memory parameters. Consequently, Edit-TA [102] mitigates the biases learnt in pre-training from a task algorithm perspective. An intuition behind it is that parameter variations learnt through negative example tasks could be perceived through weight variances. However as this fails to take the importance of different negative examples into account, therefore EWR [103] proposes Fisher information matrices to measure the uncertainty of their estimation, which is applied for the dialogue systems to execute a parameter interpolation and remove hallucination. EasyEdit [109] summarises methods for parameter editing, while minimising the influence to irrelevant parameter. An efficient alternative is to identify task-specific parameters and exploit them. For example, ALLM [106] aligns the parameter module with task-specific knowledge, and then generates the relevant knowledge as additional context in background augmented prompts. Similarly, mmT5 [55] utilises language-specific modules during pre-training to separate language-specific information from languageindependent information, demonstrating that adding language-specific modules can alleviate the curse of multilinguality. Instead, TRAC [107] combines conformal prediction and global testing to augment retrieval-based QA. The conservative strategy formulation ensures that a semantically equivalent answer to the truthful answer is included in the prediction set.

Another parameter adaptation idea focuses on flexible sampling consistent with user requirements. For instance, [51] observes that the randomness of sampling is more detrimental to factuality when generating the latter part of a sentence. The factual-nucleus sampling algorithm is introduced to keep the faithfulness of the generation while ensuring the quality and diversity. Besides, Inference-Time [108] firstly identifies a set of attentional heads with high linear probing accuracy, and then shifts activation in the inference process along the direction associated with factual knowledge.

• Post-hoc Attribution and Edit Technology. A source of hallucination is that LLMs may leverage the patterns observed in the pre-training data for inference in a novel form. Recently, ORCA [112] reveals problematic patterns in the behaviour of models by probing supporting data evidences from pre-training data. Similarly, TRAK [114] and Data-Portraits [115] analyse whether models plagiarise or reference existing resources by means of data attribution. QUIP [118] further demonstrates that providing text that has been observed in the pre-training phase can improve the ability of LLMs to generate more factual information. Furthermore, motivated by the gap between LLMs and human modes of thinking, one intuition is to align the two modes of reasoning. Thus CoT [14] elicits faithful reasoning via a kind of Chain-of-Thought (CoT) [13] prompts. Similarly, RR [113] retrieves relevant external knowledge based on decomposed reasoning steps obtained from a CoT prompt. Since LLMs do not produce the best output on the first attempt, Self-Refine [116] implements self-refinement algorithms through iterative feedback and improvement. Reflexion [117] also employs verbal reinforcement to generate reflective feedback by learning about prior failings. Verify-and-Edit [119] proposes a CoT-prompted verify-and-edit framework, which improves the fidelity of predictions by post-editing the inference chain based on externally retrieved knowledge. CoVe [120] emphasises the importance of independent self-verification to prevent being influenced by other responses. Another source of hallucinations is to describe factual content with incorrect retrievals. To recify this, NP-Hunter [111] follows a generatethen-refine strategy whereby a generated response is amended using the KG so that the dialogue system is able to correct potential hallucinations by querying the KG.

• Leverage External Knowledge. As an attempt to extend the language model for halucination mitigation, a suggestion is to retrieve relevant documents from large textual databases. RETRO [121] splits the input sequence into chunks and retrieves similar documents, while In-Context RALM [62] places the selected document before the input text to improve the prediction. Furthermore, IRCoT [122] interweaves CoT generation and document retrieval steps to guide LLMs. LLM-AUGMENTER [123] also bases the responses of LLMs on integrated external knowledge and automated feedback to improve the truthfulness score of the answers. Another work, CoK [126] iteratively analyses future content of upcoming sentences, and then applies them as a query to retrieve relevant documents for the purposes of re-generating sentences when they contain low confidence tokens. Similarly, RETA-LLM [129] creates a complete pipeline to assist users in building their own domain-based LLM retrieval systems. Note that in addition to document retrieval, diverse external knowledge queries coule be assembled into retrievalaugmented LLM systems. For example, FLARE [127] leverages structured knowledge bases to support complex queries and provide more straightforward factual statements. Further, KnowledGPT [130] adopts program of thoughts (PoT) prompting, which generates codes to interact with knowledge bases. While cTBL [125] proposes to enhance LLMs with tabular data in conversation settings. Besides, GeneGPT [124] demonstrates that expertise can be accessed more easily and accurately by detecting and executing API calls through contextual learning and augmented decoding algorithms. To support potentially millions of ever-changing APIs, Gorilla [128] explores self-instruct fine-tuning and retrieval for efficient API exploitation.

• Assessment Feedback. As language models become more sophisticated, evaluation feedback can significantly improve the quality of generated text, as well as reduce the appearance of hallucinations. To realise this concept, LSHF [131],TLM [132] and Chain-of-Hindsight [134] predict human preferences through reinforcement learning and employs this as the reward function. In addition to enabling the model to learn directly from the feedback of factual metrics in a sample-efficient manner [138], it is also important to build in a self-evaluation function of the model to filter candidate generated texts. For example, BRIO [133] empowers summarization model assessment, estimating probability distributions of candidate outputs to rate the quality of candidate summaries. While LM-know [76] is devoted to investigating whether LLMs can evaluate the validity of their own claims by detecting the probability that they know the answer to a question. Consequently, Do-LLM-Know [78] queries exclusively with black-box LLMs, and the results of queries repeatedly generated multiple times are compared with each other to pass consistency checks. As missing citation quality evaluation affects the final performance, ALCE [137] employs a natural language reasoning model to measure citation quality and extends the integrated retrieval system. Similarly, CRITIC [135] proposes to interact with appropriate tools to assess certain aspects of the text, and then to modify the output based on the feedback obtained during the verification process. Note that automated error checking can also utilise LLMs to generate text that conforms to tool interfaces. PaD [136] distills the LLMs with a synthetic inference procedure, and the synthesis program obtained can be automatically compiled and executed by an explainer. Further, iterative refinement processes are validated to effectively identify appropriate details [96,45,15], and can stop early invalid reasoning chains, beneficially reducing the phenomenon of hallucination snowballing [27].

• Mindset Society. Human intelligence thrives on the concept of cognitive synergy, where collaboration between different cognitive processes produces better results than isolated individual cognitive processes. "Society of minds" [145] is believed to have the potential to significantly improve the performance of LLMs and pave the way for consistency in language production and comprehension. For the purpose of addressing hallucinations in large-scale multilingual models across different translation scenarios, HLMTM [39] proposes a hybrid setting in which other translation systems can be requested to act as a back-up system when the original system is hallucinating. Consequently, Multiagent-Debate [140] employs multiple LLMs in several rounds to propose and debate their individual responses and reasoning processes to reach a consensus final answer. As a result of this process, the models are encouraged to construct answers that are consistent with both internal criticisations and responses from other agents. Before a final answer is presented, the resultant community of models can hold and maintain multiple reasoning chains and possible answers simultaneously. Based on this idea, MAD [141] adds a judge-managed debate process, demonstrating that adaptive interruptions of debate and controlled "titfor-tat" states help to complete factual debates. Furthermore, FORD [142] proposes roundtable debates that include more than two LLMs and emphasises that competent judges are essential to dominate the debate. LM-vs-LM [80] also proposes multi-round interactions between LM and another LM to check the factualness of original statements. Besides, PRD [143] proposes a peer rank and discussionbased evaluation framework to arrive at a well-recognised assessment result that all peers are in agreement with. In an effort to maintain strong reasoning, SPP [144] utilises LLMs to assign several fine-grained roles, which effectively stimulates knowledge acquisition and reduces hallucinations.

Future Directions

Though numerous technical solutions have been proposed in the survey for hallucinations in LLMs, there exist some potential directions:

• Data Construction Management. As previously discussed, the style, and knowledge of LLMs is basically learned during model pre-training. High quality data present promising opportunities for facilitating the reduction of hallucinations in LLMs [146]. Inspired by the basic rule of machine learning models "Garbage input, garbage output", [147] proposes that data quality and diversity outweigh the importance of fine-tuning large-scale instructions [148,3,149] and RLHF [6,2]. To perform efficiently in knowledge-intensive verticals, we argue that construction of entity-centred fine-tuned instructions [150,151,152] is a promising direction that it can enhance the factuality of generated entity information. Another feasible proposal is to incorporate a self-curation phase [153] in the instruction construction process to rate the quality of candidate pairs. During the iteration process, quality evaluation [154] based on manual or automated rule constraints could provide self-correction capacity.

• Reasoning Mechanism Exploitation. The emerging CoT technique [14] stimulates the emergent reasoning ability of LLMs by imitating intrinsic stream of thought. Recently, A primary improvement is ToT [155] introduces tree and into the thought process, and provides a novel backtrack function. However, the actual thinking process creates a complex network of ideas, as an example, people could explore a particular chain of reasoning, backtrack or start a new chain of reasoning. GoT [156] extends the dependencies between thoughts by constructing vertices with multiple incoming edges to aggregate arbitrary thoughts. Since previous methods have no storages for intermediate results, CR [156] uses cumulative and iterative manners to simulate human thought processes, and decompose the task into smaller components. In addition to self-heuristic methods, PAL [157] and PoT [158] introduce programming logic into the language space [159], expanding the ability to invoke external explainers. As a summary, research based on human cognition helps to provide brilliant insights into the analysis of hallucinations, such as Dual Process Theory [160], Three layer mental model [161], Computational Theory of Mind [162], and Connectionism [163].

• Multi-modal Hallucination Survey. It has become a community consensus to establish powerful Multimodal Large Language Models (MLLMs) [164,165,166] by taking advantage of excellent comprehension and reasoning capabilities of LLMs. [41] confirms the severity of hallucinations in MLLM by object detecting and polling-based querying. The results indicate that they are highly susceptible to object hallucination, and the generated description does not match the target image. Besides, [167] that MLLMs have limited multimodal reasoning ability as well as dependence on spurious cues. Though current study [168] provides a broad overview of MLLMs, the causation of hallucinations has not been comprehensively investigated. In the future, as more sophisticated multi-model applications emerge, in-depth analyses of the biased distribution resulting from misalignment among modes is a promising research direction, to provide faithful modal interactions.

Conclusion and Vision

In this paper, we provide an overview of hallucinations in LLMs with new taxonomy, theoretical insight, detection methods, correction methods and several future research directions. Note that it is crucial to utilize LLMs in a responsible and beneficial manner. Furthermore, with sophisticated and efficient detection methods proposed for various aspects, LLMs will provide human reliable and secure information in broad application scenarios.