Hallucination of Multimodal Large Language Models: A Survey

Introduction

• MLLM hallucination research primarily concerns cross-modal inconsistency between generated responses and visual content, unlike pure-LLM factuality and faithfulness hallucinations.Factuality concerns divergence from real-world facts; faithfulness concerns divergence from instructions or context and self-consistency.Source: §1 · p. 2
• The survey classifies object hallucinations as category, attribute, or relation errors, grouping object counting and events under attributes.Categories include nonexistent or incorrect object classes, inaccurate object properties, and incorrect relationships among objects.Source: §1 · p. 2
• The survey analyzes MLLM-specific causes across data, models, training, and inference, then reviews dedicated evaluation benchmarks, metrics, and cause-linked mitigation strategies.Its mitigations are organized around the identified causes.Source: §1 · pp. 2–3
• The survey aims to clarify current MLLM limitations and guide future research toward more robust and trustworthy models.It closes by examining challenges and open questions that frame current limitations and future prospects.Source: §1 · pp. 2–3
• Compared with pure-LLM hallucination surveys [65] [209] and the concurrent short LVLM survey [112], this work offers a broader literature scope and layered taxonomy beyond specific architectures.It also traces causes across affecting factors and links mitigation strategies to those causes.Source: §1 · pp. 2–3

2.1 Large Language Models

• LLMs are transformer-based models trained on vast textual datasets, with scaling associated with emergent in-context learning, chain-of-thought prompting, and instruction following.Examples of these capabilities are In-Context Learning [14], Chain-of-Thought prompting [168], and instruction following [130].Source: §2.1 · p. 3
• LLMs typically undergo pre-training, supervised fine-tuning, and reinforcement learning from human feedback to develop language capabilities and align outputs with instructions and preferences.Pre-training uses autoregressive prediction on text; SFT uses annotated instruction-response pairs; RLHF uses a preference model to reward outputs, often with PPO [143].Source: §2.1 · pp. 3, 5

2.2 Multimodal Large Language Models

• MLLMs enable LLMs to perceive visual data, supporting tasks such as captioning and visual question answering; this survey focuses on visual-to-text understanding.MLLMs are also called VLLMs or LVLMs, and the survey distinguishes its understanding focus from unified models that combine understanding and generation.Source: §2.2 · p. 5
• MLLMs connect vision and language either through an interface between pretrained unimodal models or through end-to-end training from scratch.Interfaces use learnable queries with cross-attention or projection layers to map visual features into the LLM input space; Fuyu-8B instead projects raw image-patch pixels into embeddings without a pretrained vision encoder.Source: §2.2 · pp. 5–6
• Interface-based MLLMs typically use pre-training for cross-modal feature alignment and instruction tuning to improve multimodal instruction following.Pre-training commonly freezes the visual encoder and LLM while training the interface on image-text pairs; instruction tuning uses visual-instruction-response data, whose quality affects performance.Source: §2.2 · p. 6

2.3 Hallucinations in Multimodal Large Language Models

• Object hallucination occurs when generated text conflicts with visual content, commonly through category, attribute, or relation errors.The survey prioritizes incorrect descriptions over missing objects because correctness is more important to human judges than specificity, while object coverage is less perceptible. Counting and events are classified as attributes; verb hallucination is not treated independently.Source: §2.3 · pp. 6–7, Figure 3 · p. 7

3 HALLUCINATION CAUSES

• Hallucination causes span MLLMs’ capability acquisition process and are grouped into data, model, training, and inference factors.Source: §3 · p. 7

3.1 Data

• Data enables cross-modal understanding and instruction following, but can also cause hallucinations through quantity, quality, and statistical bias.Source: §3.1 · p. 7

3.1.1 Quantity.

• Image-text and visual-QA datasets remain far smaller than text-only LLM training data, and insufficient data can impair cross-modal alignment and cause hallucinations.Source: §3.1.1 · p. 7

3.1.2 Quality.

• Inaccurate, misaligned, or corrupted pretraining pairs limit cross-modal alignment; GPT-4-generated instructions risk noise, while LLaVA-1.5’s human-annotated QA improved results.Source: §3.1.2 · pp. 7–8
• Instruction tuning often lacks negative instructions and reject answers, encouraging models to answer “Yes” even when “No” is appropriate and thereby hallucinate.Source: §3.1.2 · p. 8
• Description detail remains an open question: sparse object-position, attribute, and non-salient-object descriptions may weaken grounding, while overly detailed instructions may prompt unsupported details.The survey presents both insufficient detail and detail exceeding MLLMs’ perceptual limits as possible risks.Source: §3.1.2 · p. 8

3.1.3 Statistic bias.

• Training-set frequency and co-occurrence biases can make models predict absent common objects or infer associated objects; scaling data may alleviate, but not eliminate, bias.Source: §3.1.3 · p. 8

• Weak vision models can misclassify or misinterpret visual concepts, and information loss during encoding can undermine multimodal understanding [137] [160].Source: §3.2 · p. 9
• Because language models are often stronger than vision models, parametric knowledge can override visual content and cause hallucinations [93] [94] [109] [137].Source: §3.2 · p. 9
• Weak alignment interfaces can cause hallucinations through data, architecture, or training-loss limitations [75] [117].The LLaVA-like linear projection preserves information but lacks projected-feature supervision, leaving a distribution gap [75]; Q-former-like alignment uses diverse supervision but learnable queries can lose fine-grained visual information [98].Source: §3.2 · p. 9

3.3 Training

• MLLMs generally train with the same autoregressive next-token prediction loss used for LLMs.The passage describes this loss as straightforward, effective, and easy to scale for language modeling.Source: §3.3 · p. 9
• Some studies question next-token loss for visual learning because visual content has complex spatial structure and the loss lacks sequence-level supervision [11] [29].Source: §3.3 · p. 9
• The absence of an RLHF stage in MLLM training has been proposed as a potential cause of hallucination [148] [193].Source: §3.3 · p. 9

3.4 Inference

• As autoregressive sequences grow, attention shifts toward previously generated text and special tokens rather than visual content, making responses irrelevant to the image [41] [66] [157] [160] [161].Attention-map visualizations show increased focus on previously generated tokens such as punctuation [66] [161].Source: §3.4.1 · p. 9

3.4.1 Visual Attention Deficiency.

• AvisC [170] finds that a few visual tokens receive excessive attention while lower-weight tokens may encode nuanced object details, impairing fine-grained recognition; VTI links noise-sensitive vision tokens to representation outliers and hallucinations.Source: §3.4.2 · p. 10

3.4.2 Trap Visual Tokens.

• Recent work [19] [21] finds that a small portion of visual tokens can strongly affect final results and induce hallucinations.Source: §3.4.2 · p. 10

Core object-hallucination metrics and benchmarks

• CHAIR [137] measures object hallucination in captions, while POPE [103] and MME [187] use yes-or-no object questions.CHAIR has instance- and sentence-level variants and remains widely used. POPE adds positive and negative questions with random, popular, and adversarial negative sampling; MME tests object existence, count, position, and color.Source: §4 · pp. 10–11, Table 1 · p. 10

LLM-assisted and generated evaluations

• CIEM [63], MMHal-Bench [148], GAVIE [109], and HaELM [161] broaden evaluation through generated data, open-ended judging, and trained evaluators.CIEM automatically generates QA pairs, while MMHal-Bench uses GPT-4 to rate responses. GAVIE uses GPT-4 to score relevance and accuracy without human-annotated answers; HaELM trains an LLM to evaluate hallucinations against image descriptions.Source: §4 · pp. 11–12

Fine-grained and diagnostic evaluations

• NOPE [117], FaithScore [79], AMBER [160], and Bingo [35] target answerability, fine-grained facts, multidimensional tasks, and bias or interference.NOPE distinguishes hallucination from incorrectness and supplements scarce negative-answer data. FaithScore decomposes responses into facts for verification; AMBER combines CHAIR and F1, while Bingo analyzes hallucination patterns using success and failure cases.Source: §4 · p. 12

Expanded benchmark dimensions

• CCEval, RAH-Bench [29], HallusionBench [108], MERLIM [154], FGHE [163], and OpenCHAIR [11] extend evaluation to detail, relations, reasoning, and open vocabulary.CCEval adds caption coverage; RAH-Bench probes categorical, attribute, and relation hallucinations. HallusionBench separates visual-dependent from visual-supplement questions, MERLIM uses edited image pairs, FGHE tests fine-grained binary questions, and OpenCHAIR relaxes CHAIR’s closed vocabulary.Source: §4 · pp. 12–13

Additional hallucination types

• Hal-Eval [74], CorrelationQA [55], and VQAv2-IDK [17] address event, spurious-visual-input, and answerability-related hallucinations.CorrelationQA finds hallucinations in responses to relevant but inconsistent images. VQAv2-IDK uses unanswerable image-question pairs and categorizes “I Know” hallucinations into four types.Source: §4 · pp. 13–14

Detection and evaluation frameworks

• MHaluBench [25], VHTest [67], THRONE [82], and BEAF [186] assess detection tools, visual-property modes, free-form outputs, and scene-change understanding.MHaluBench evaluates hallucination-detection methods rather than MLLMs themselves. VHTest supports open-ended and yes-or-no formats; THRONE targets free-form outputs with an open-source LLM, while BEAF evaluates perception through edited scenes and introduces four metrics.Source: §4 · pp. 14–15

Comparison of mainstream models

• MLLM performance is not consistent across benchmarks, whose evaluation dimensions and emphases differ.The survey compares models on generative and discriminative tasks; Table 2 covers generative benchmarks.Source: Table 2 · p. 15, §4 · p. 15

5 HALLUCINATION MITIGATION

• Mitigation approaches are categorized by perspective into Data, Model, Training, and Inference methods.The section presents a comprehensive review of contemporary methods aimed at mitigating MLLM hallucinations.Source: §5 · p. 15

5.1 Data

• Data-focused mitigation introduces negative, counterfactual, and reasoning data while reducing noise and errors in existing datasets.The survey identifies data as a primary factor inducing hallucination in MLLMs.Source: §5.1 · p. 15

5.1.1 Negative Data.

• LRV-Instruction [109] counters positive-only tuning that encourages “Yes” answers with negative examples manipulating nonexistent objects, existent-object attributes, and instruction knowledge.HalluciDoctor [191] detects hallucinated content through crosschecking multiple MLLMs, removes it, and uses counterfactual generation to address long-tail distributions and object co-occurrence; models trained on its calibrated data are less prone to hallucination.Source: §5.1.1 · pp. 15–16, §5.1.2 · p. 16

5.1.2 Counterfactual Data.

• PerturboLLaVA [22] trains with text perturbations that align with general knowledge but conflict with the image, encouraging image-grounded predictions over language-prior-driven hallucinations.The approach shifts the model’s conditional distribution toward the image and away from perturbation text, yielding more robust multimodal capability.Source: §5.1.2 · pp. 16–17
• REVERIE [202] uses reflective instruction tuning to train models to predict rationales explaining why responses are correct or incorrect, rather than learning from responses alone.Its curated examples pair correct and confusing responses with rationales justifying their correctness or errors.Source: §5.1.3 · p. 17

5.1.3 Reasoning Data.

• ReCaption [163] rewrites captions by extracting keywords and using an LLM to generate sentences, producing image-caption pairs whose trained models achieve higher accuracy on benchmarks such as POPE [103].Why rewritten captions can reduce hallucination remains an open question.Source: §5.1.4 · p. 17

5.1.4 Clean Data.

• EOS Decision [196] filters harmful training data and strengthens timely sequence termination to reduce hallucinations caused by captions demanding details beyond the model’s perception limit.The method is motivated by the observation that hallucinations tend to occur for objects mentioned later in generated descriptions [224].Source: §5.1.4 · p. 17

5.2.1 Scale-up Resolution.

• Higher vision resolution generally lowers hallucination: on CCEval, HallE-Switch found lower hallucination across CLIP-ViT-L-112, -224, and -336; LLaVA-1.5 [110] scales its CLIP ViT encoder from 224 to 336.Source: §5.2.1 · p. 17

5.2.2 Versatile Vision Encoders.

• CLIP’s loss of visual detail motivates combining it with DINO ViT [16], but additive and interleaved feature mixing each show trade-offs [152].A visual expert model [59] instead integrates latent information through multitask encoders and uses OCR and object detectors for structural knowledge enhancement.Source: §5.2.2 · pp. 17–18
• VCoder [71] adds segmentation masks and depth maps, another model [77] integrates object detection and OCR through several training strategies, and EAGLE [155] retrains the encoder for visual grounding.The described approaches enhance perception or object identification to address hallucination.Source: §5.2.2 · p. 18
• HallE-Switch trains a switch module on contextual and parametric data to control parametric knowledge in detailed captions by tuning 𝜖 at inference.The method targets hallucinations that arise when language knowledge guides generation instead of visual content.Source: §5.2.3 · p. 18

5.2.3 Dedicated Module.

• PATCH [144] inserts trainable virtual tokens between image features and enhanced prompts to align visual and textual features and mitigate object hallucinations.Fine-tuned token embeddings are added to the vocabulary at inference, making the method plug-and-play.Source: §5.2.3 · p. 18

5.3.1 Auxiliary supervision.

• RAI-30k adds panoptic-scene-graph relation annotations and SAM-based mask prediction loss to instruction tuning, encouraging focus on relevant instances and mitigating vision hallucination [29].The survey presents grounding supervision as promising for reducing hallucination.Source: §5.3.1 · pp. 18–19
• HACL [75] uses hallucinated text as hard negatives for visual anchors; HALVA [140] aligns correct and altered response phrases, while VISTA [104] reinforces visual cues and meaningful tokens.HACL contrasts visual and textual embeddings; HALVA constructs correct/hallucinated response pairs, and VISTA uses positive and negative prompts.Source: §5.3.1 · p. 19
• Selective EOS Supervision modifies MLE to teach proper termination [196], while CCA [178] reduces visual-to-instruction token distance to counter RoPE long-term decay.CCA addresses the observed tendency for hallucination to increase when relevant visual cues are distant from instruction tokens.Source: §5.3.1 · p. 19

5.3.2 Reinforcement Learning.

• MOCHa [11] optimizes image-captioning rewards for NLI fidelity, BERTScore [206] semantic adequacy, and KL regularization using proximal policy optimization.Its multi-objective reward seeks accurate, relevant, and varied captions while keeping the policy close to its initial version.Source: §5.3.2 · pp. 19–20
• AI-feedback preference training constructs accurate and hallucinatory pairs for DPO, with later work adding fine-grained feedback, vision-guided pairs, or on-policy expert corrections [214] [101] [223] [194] [78] [183].Examples include GPT-4-based detection or preference distillation, AI-generated dispreferences, and object-, attribute-, relation-, or segment-level feedback.Source: §5.3.2 · pp. 20–21
• Human-feedback approaches train reward models or directly optimize fine-grained preferences to improve faithfulness, including FDPO [52], Factually Augmented RLHF [148], DDPO [193], and ViGoR [180].The methods use hallucination labels, human comparisons or segment-level corrections, and—in ViGoR—automatic grounding and fidelity scores.Source: §5.3.2 · p. 21
• Visual generative feedback uses caption-reconstructed images to guide PPO or decoding [128], while DeGF [201] uses generated images as auxiliary references for self-correction.DeGF applies complementary or contrastive decoding and is described as training-free.Source: §5.3.2 · pp. 21–22
• EFUF [177] applies sub-sentence positive and hallucinated-sample losses using CLIP-constructed data, introducing unlearning as a path for hallucination mitigation.The survey characterizes it as the first and, at the time described, only MLLM hallucination-mitigation work using unlearning.Source: §5.3.3 · p. 22

5.4.1 Generation Intervention.

• Contrastive decoding calibrates outputs against a distorted-input distribution to reduce visual-bias and language-prior reliance, with later methods varying how they induce and use the contrast.VCD contrasts original and distorted visual inputs; subsequent approaches disturb prompts, tokens, or visual features, or use auxiliary models and image augmentations.Source: §5.4.1 · pp. 22–23
• Other decoding interventions guide generation with grounded visual evidence, amplify image information, or reuse earlier-layer knowledge to correct hallucination-prone outputs.MARINE grounds objects, HALC finds token-wise visual contexts, M3ID amplifies the image-conditioned distribution, and DeCo uses preceding-layer knowledge to correct output logits.Source: §5.4.1 · pp. 23–24

5.4.2 Visual Prompting.

• Visual prompting methods perturb or augment visual inputs to steer responses toward image content, with one study also examining background information’s role in object hallucination.VAP optimizes visual perturbations to redirect decisions away from parametric bias; SoM-LLaVA applies Set-of-Mask prompting to reduce hallucination.Source: §5.4.2 · pp. 25–26

5.4.3 RAG.

• Retrieval augmentation targets hallucinations by retrieving trustworthy external information selectively, including when the model has low certainty or lacks relevant knowledge.ARA decomposes the hallucinated target object, selects an appropriate retrieval method, and triggers retrieval selectively to reduce cost.Source: §5.4.3 · p. 26

5.4.4 Ensembling.

• Ensembling approaches combine transformed-image predictions, multi-agent debate, or multi-view and multi-path reasoning; Dentist first classifies queries to select mitigation strategies.MVP aggregates certainty across answer paths, while RITUAL uses random image transformations to improve robustness.Source: §5.4.4 · p. 26

5.4.5 Post-hoc Correction.

• Post-hoc correction detects and revises hallucinated content through visual grounding, trained revisors, self-revision, logical consistency checks, or tool-assisted verification.Woodpecker validates extracted claims against visual evidence; LURE trains a revisor, Volcano iteratively revises responses, LogicCheckGPT checks attribute-object loops, and VFC verifies proposed captions with tools.Source: §5.4.5 · pp. 26–27

6.1 Data-centric Challenges and Innovations

• Training reliable MLLMs requires high-quality, diverse data, with future work on data collection, augmentation, and recalibration to reduce bias and promote inclusivity.The survey identifies data scarcity and diversity as concerns, and calls for recalibrating existing datasets to mitigate issues that may induce hallucinations.Source: §6.1 · pp. 27–28

6.2 Cross-modal Alignment and Consistency

• Cross-modal alignment requires better modality representations and supervision for training, alongside unified consistency-evaluation approaches drawing on multimodal understanding and text-to-image research [75] [29] [98] [25] [30].The survey proposes advanced architectures and learning objectives [75], diverse supervision signals [29], and either adapting existing consistency checks or unifying them into an overall framework.Source: §6.2 · p. 28

6.3 Advancements in Model Architecture

• Hallucination-focused architectures could strengthen visual perception, improve cross-modal information transfer, and make language models more faithful to visual inputs and text instructions.The survey recommends designing these architectures based on identified causes of hallucination.Source: §6.3 · p. 28

6.4 Establishing Standardized Benchmarks

• Standardized, theoretically sound, easy-to-use benchmarks are needed because binary Yes-or-No tests may not reflect human use and generative evaluations often depend on external models.Without sound benchmarks, hallucination-mitigation research may be built on an incorrect foundation.Source: §6.4 · p. 28

6.5 Reframing Hallucination as a Feature

• Hallucination could be treated as a feature when it enriches user experience or inspires creative ideas, shifting optimization beyond specific cross-modal benchmarks.The survey distinguishes content that causes hallucinations but does not affect user experience from content that does.Source: §6.5 · p. 29

6.6 Enhancing Interpretability and Trust

• Because mitigation methods often target empirically observed patterns while decision processes remain unclear, research should interpret generation and trace how inputs lead to outputs.Suggested work includes visualizing model internals and identifying salient features and linguistic patterns, which could help users assess hallucinated content.Source: §6.6 · p. 29

6.7 Navigating the Ethical Landscape

• Responsible MLLM development should address misinformation, bias, privacy, and societal impact across the lifecycle, from data collection and training through deployment and evaluation.The survey argues that ethical considerations should accompany work on hallucinations beyond typical object hallucination.Source: §6.7 · p. 29

7 CONCLUSION

• The survey reviews MLLM hallucination causes, evaluation metrics and benchmarks, and mitigation methods.Source: §7 · p. 29
• Hallucination remains a complex, persistent concern that threatens the practical application, safety, reliability, and trustworthiness of MLLMs.Source: §7 · p. 29
• The survey outlines future research directions and aims to serve as a foundational resource for researchers and practitioners developing robust solutions.Source: §7 · p. 29