Research map / Vision-language and multimodal / Multimodal LLMs
HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces HallusionBench, a human-expert-crafted benchmark of 346 images and 1129 yes/no questions organized into control pairs (original vs. human-edited images, with or without visual input) to diagnose language hallucination and visual illusion in LVLMs, evaluating 15 models with GPT-4-assisted judging.
Problem
Existing VLM benchmarks focus on object hallucination and rarely diagnose whether LVLM failures come from over-reliance on language priors (language hallucination) or from misperceiving visual input (visual illusion).
Approach
Pair each question with a control counterpart on an original and a human-edited image (or no image), so answer consistency and correctness across the pair reveal whether a failure stems from parametric priors or visual misinterpretation; a decision tree labels failure types.
Key results
- GPT-4V achieves the top question-pair accuracy of 31.42% (human evaluation) and 28.79% (GPT-4-assisted), while all other evaluated models fall below 16% qAcc.
- LLaVA-1.5 (13B) scores 10.55% qAcc under GPT-4-assisted judging, and its yes-bias analysis shows it tends to answer 'yes'.
- GPT-4-assisted judging closely matches human judgment (GPT-4V qAcc 28.79 vs 31.42 human), supporting the automatic evaluation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPT-4V instead of the other evaluated LVLMs | LVLM evaluation on HallusionBench · HallusionBench | Question Pair Accuracy (qAcc): 28.79 vs. all other evaluated models below 16% | Helped |
| Evaluate on human-edited images rather than the easier questions | GPT-4V and other open-source LVLMs · HallusionBench Hard | Accuracy: GPT-4V: 75.60 -> 37.67 | Hurt |
| Increase the LLM backbone size | Open-source LVLMs · HallusionBench | All Accuracy (aAcc) | Helped |
| Use broader instruction fine-tuning across 26 diverse datasets | InstructBLIP · HallusionBench | All Accuracy (aAcc) | Helped |
| Use the language decoder as a universal interface for managing modalities | mPLUG-Owl-v2 compared with mPLUG-Owl-v1 · HallusionBench | All Accuracy (aAcc): 43.93 -> 47.30 | Helped |
| Evaluate with GPT-4-assisted scoring rather than human evaluation | GPT-4V · HallusionBench | Question Pair Accuracy (qAcc): 31.42 -> 28.79 | No clear effect |
Surprising findings
- GPT-4V can recognize famous optical illusions by name yet fails to answer correctly on human-edited versions of the same images, suggesting reliance on memorized priors.
- Several models perform worse than random chance on question-pair accuracy, including Qwen-VL and MiniGPT-4.
Limitations
- Relies on GPT-4 as a text-only judge, with residual output variance even at temperature 0.
- Benchmark size is modest (346 images, 1129 questions) and focused on yes/no questions.
- Failure-type diagnosis is rule-based via a decision tree and may label mixed cases as 'potentially mixed'.
Experimental setup
Benchmarks: HallusionBench
Metrics: question pair accuracy (qAcc) figure accuracy (fAcc) all accuracy (aAcc) yes percentage difference false positive ratio consistency language hallucination vs visual illusion rate
Baselines: GPT-4V LLaVA-1.5 Gemini Pro Vision Claude 3 BLIP2-T5 Qwen-VL InstructBLIP mPLUG-Owl-v2 Random Chance
Builds on: POPE MME GAVIE Bingo
Closest papers by meaning
- MIRAGE: Assessing Hallucination in Multimodal Reasoning Chains of MLLM (NeurIPS 2025 · similarity 0.92)
- AVHBench: A Cross-Modal Hallucination Benchmark for Audio-Visual Large Language Models (ICLR 2025 · similarity 0.92)
- Fine-Grained Multi Image Object Hallucination Benchmark (CVPR 2026 · similarity 0.92)
- Analyzing and Mitigating Object Hallucination in Large Vision-Language Models (ICLR 2024 · similarity 0.91)
- Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.91)
- Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens (CVPR 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.