| task_categories: | |
| - image-text-to-text | |
| # MVI-Bench | |
| MVI-Bench is a comprehensive benchmark specifically designed to evaluate the robustness of Large Vision-Language Models (LVLMs) against misleading visual inputs. | |
| - **Paper:** [MVI-Bench: A Comprehensive Benchmark for Evaluating Robustness to Misleading Visual Inputs in LVLMs](https://huggingface.co/papers/2511.14159) | |
| - **Repository:** [https://github.com/chenyil6/MVI-Bench](https://github.com/chenyil6/MVI-Bench) | |
| ## Introduction | |
| Grounded in fundamental visual primitives, the design of MVI-Bench centers on three hierarchical levels of misleading visual inputs: | |
| 1. **Visual Concept** | |
| 2. **Visual Attribute** | |
| 3. **Visual Relationship** | |
| Using this taxonomy, the benchmark compiles 1,248 expertly annotated VQA instances across six representative categories. It also introduces **MVI-Sensitivity**, a novel metric that characterizes LVLM robustness at a granular level. Empirical results across 18 state-of-the-art LVLMs uncover pronounced vulnerabilities to misleading visual inputs. | |
| ## Dataset Structure | |
| The benchmark evaluates how misleading visual information affects model understanding and decision-making in visual question-answering tasks. It covers 1,248 instances designed to test models against various visual misinformation scenarios across different levels of complexity. |