Source-linked AI summary
A Survey on Hallucination in Large Vision-Language Models
Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, Wei Peng
TL;DR
LVLM hallucinations—misalignments between visual facts and generated text—challenge practical deployment. This survey synthesizes their symptoms, evaluation benchmarks, causes, mitigation methods, and future directions, identifying data, vision encoders, modality alignment, and output processing as central topics. It concludes that these analyses provide a foundation for more practical LVLM research while noting constraints in visual resolution, token capacity, and resource costs.
Problem
Hallucinations misalign LVLM-generated text with factual visual content, creating a significant challenge for practical LVLM application.
Method
The survey reviews hallucination concepts, evaluation methods and benchmarks, causes in training data and model components, mitigation approaches, and future directions.
Results
The survey identifies biased training data, vision-encoder grounding limits, modality misalignment, insufficient context attention, and generative language-model behavior among hallucination causes, and reviews data, module, and post-processing mitigation strategies.
Takeaways & Limitations
The synthesis aims to guide future research and development toward more reliable and efficient LVLMs for practical implementation.
Takeaways & Limitations
Restricted visual-token quantities can prevent alignment modules from encoding all image information, increasing hallucination risk.
Abstract
from arXiv · showhide
Recent development of Large Vision-Language Models (LVLMs) has attracted growing attention within the AI landscape for its practical implementation potential. However, ``hallucination'', or more specifically, the misalignment between factual visual content and corresponding textual generation, poses a significant challenge of utilizing LVLMs. In this comprehensive survey, we dissect LVLM-related hallucinations in an attempt to establish an overview and facilitate future mitigation. Our scrutiny starts with a clarification of the concept of hallucinations in LVLMs, presenting a variety of hallucination symptoms and highlighting the unique challenges inherent in LVLM hallucinations. Subsequently, we outline the benchmarks and methodologies tailored specifically for evaluating hallucinations unique to LVLMs. Additionally, we delve into an investigation of the root causes of these hallucinations, encompassing insights from the training data and model components. We also critically review existing methods for mitigating hallucinations. The open questions and future directions pertaining to hallucinations within LVLMs are discussed to conclude this survey.
1 Introduction
LVLMs extend language-model capabilities to vision-language tasks, but hallucinations—disagreements between image facts and generated text—impede practical use. This survey characterizes symptoms, causes, mitigation methods, and future research directions.
- LVLMs insert visual features into language models and align them with textual features for vision-language tasks.
- LVLM hallucinations include errors in judgment and description, spanning objects, attributes, and relations.
- Hallucinations denote disagreements between factual image content and corresponding generated textual content.
- Unlike captioning studies centered mainly on object presence, LVLMs produce more detailed descriptions that diversify hallucination symptoms.
- The survey analyzes hallucination causes, mitigation methods, and future directions to support more reliable and efficient LVLM development.
2 Hallucination in the Era of LVLM
LVLM hallucinations are contradictions between visual facts and textual outputs, appearing as judgment or description errors across multiple semantic dimensions. Their causes and mitigation are intertwined with data quality, visual uncertainty, representation limits, and multimodal alignment.
- LVLM Architecture and Training: LVLMs combine a visual encoder, modality connection module, and language model, trained through pre-training and instruction-tuning.
- Hallucination Definition and Symptoms: Hallucinations contradict visual input and textual output, appearing as judgment failures or unfaithful descriptions.
- Unique Challenges: Semantic hallucinations involve nonexistent objects, incorrect attributes, or inaccurate relations, requiring fine-grained visual-text alignment for detection.
- Unique Challenges: Unclear or distorted images can exacerbate language priors and statistical biases, producing more severe hallucinations.
- Mitigation Challenges: Mitigation combines data quality enhancement, encoding optimization, visual-representation refinement, and multimodal alignment, while high-resolution encoders remain resource-demanding.
3 Evaluation Methods and Benchmarks
The survey organizes LVLM hallucination evaluation into non-hallucinatory generation and hallucination discrimination, supported by handcrafted, model-based, and question-answering methods. Benchmarks distinguish discriminative object-level testing from generative evaluations covering broader semantic errors.
- Evaluation Taxonomy: Evaluation methods assess either non-hallucinatory content generation or hallucination discrimination, with corresponding generative and discriminative benchmarks.
- Generation Evaluation: Non-hallucinatory generation measures hallucinated output content through handcrafted pipelines or model-based end-to-end methods.
- Generation Evaluation: FAITHSCORE evaluates free-form responses without references by extracting sub-sentences and atomic facts, then comparing them with the input image.
- Discrimination Evaluation: Hallucination discrimination uses questions whose descriptions agree or conflict with image content, assessing the model’s responses.
- Benchmarks: POPE, NOPE, and CIEM are discriminative benchmarks focused on object hallucinations and evaluated with accuracy, using datasets of 3000, 17983, and 72941 examples, respectively.
- Benchmarks: Generative benchmarks cover object, attribute, and relation hallucinations, while AMBER combines generative and discriminative tasks.
4 Causes of LVLM Hallucinations
LVLM hallucinations arise from deficiencies in training data, visual encoding, modality alignment, and language-model behavior. These factors include biased or irrelevant annotations, incomplete visual grounding, information loss during alignment, and decoding or attention patterns that favor fluent but inaccurate content.
- Hallucinations from Data: Biased and homogeneous training data can produce consistently affirmative answers, weak instruction following, and inaccurate descriptions of local visual relations.Factual judgment data may overrepresent “Yes” answers, while limited instruction and visual diversity restrict robust visual understanding.
- Hallucinations from Data: LLM-generated annotations may include objects, attributes, and relationships that do not match fine-grained image content, catalyzing hallucinations.
- Hallucinations from Vision Encoders: CLIP-derived vision encoders can miss fine-grained visual semantics because of restricted resolution and limited visual information expression.
- Hallucinations from Modality Aligning: Simple connection modules and limited alignment tokens can leave visual and textual features misaligned or discard image information, increasing hallucination risk.Linear layers may provide insufficient multimodal connectivity, while Q-Former-style modules encode a predetermined number of tokens rather than all image information.
- Hallucinations from LLMs: LLM-related causes include insufficient context attention, stochastic sampling, and capability misalignment beyond the model’s established knowledge.During decoding, models may neglect visual inputs, prioritize language patterns, or focus on partial generated summaries.
5 Mitigation of LVLM Hallucination
Mitigation methods target the data, visual representation, modality alignment, decoding, human-preference training, and post-processing stages of LVLM pipelines. The survey describes both training-based interventions and output-editing approaches that use visual evidence to reduce inaccuracies.
- Mitigation for Data: Optimizing training data addresses bias and annotation deficiencies through contrastive instruction tuning, diverse visual instructions, and richly grounded datasets.Examples include LRV-Instruction’s 400,000 visual instructions, M-HalDetect’s 4,000 annotated image-description pairs, and GRIT’s 1.1 million spatially grounded samples.
- Mitigation for Vision Encoders: Scaling image resolution and enriching perceptual inputs improve visual representation by supporting finer object recognition, spatial awareness, and visual-cue extraction.Approaches include patch-based high-resolution processing, larger vision encoders, and auxiliary segmentation or depth modalities.
- Mitigation for Connection Module: More capable connection modules and additional alignment objectives reduce modality gaps by improving correspondence between visual and textual features.LLaVA-1.5 replaces a linear connector with an MLP, while alignment objectives and RLHF are used to bring modalities closer and reduce hallucinations.
- Mitigation for LLMs: Decoding optimization modifies generation to focus on visual context, including OPERA’s weighted beam-search scoring that de-prioritizes over-focused summary tokens.
- Mitigation via Post-processing: Post-processing methods refine generated descriptions using visual validation, with LURE training a revisor and training-free Woodpecker correcting content through a structured visual knowledge base.
6 Future Directions
Future work targets finer supervision, richer modality coupling, tool-using LVLM agents, and interpretable hallucination mechanisms. These directions aim to improve visual detail understanding, grounding fidelity, and mitigation efficiency.
- Supervision Objective: Current coarse-grained image-caption alignment and whole-image instruction tuning limit LVLMs’ ability to understand complex visual details.Object masked language modeling and position-enhanced instruction following are proposed as more detailed, spatially aware alternatives.
- Enriching Modalities: Integrating multiple modalities can enhance model capabilities and performance on single-modal tasks through correlated and complementary information.Video incorporation through cache retrieval and prealignment has shown benefits for addressing existence-related hallucinations.
- LVLM as Agent: LVLMs can use visual tools such as detection and segmentation models to process specific visual details and reason over structured outputs.Agent-style LVLMs have demonstrated fidelity in OCR-VQA, object tagging, and grounding.
- Delving into Interpretability: Post-processing methods can address hallucinations by refining generated outputs, but they require careful design and higher computational costs.Further study of internal hallucination mechanisms could support interpretability-oriented solutions.
7 Conclusion
The survey investigates LVLM hallucinations through evaluation methods, benchmarks, causal analysis, mitigation approaches, and future directions. It aims to establish a foundation for addressing these complexities and advancing practical LVLM applications.
- 7 Conclusion: The survey investigates LVLM hallucinations and introduces evaluation methods and pertinent benchmarks.It focuses on hallucinations that challenge LVLMs’ practical application in open-domain vision-language tasks.
- 7 Conclusion: The survey analyzes fundamental causes, discusses mitigation approaches, and identifies existing challenges and potential directions.Its stated aim is to facilitate future research toward practical implementation of LVLMs.