KEBench: A Benchmark on Knowledge Editing for Large Vision-Language Models

Currently, little research has been done on knowledge editing for Large Vision-Language Models (LVLMs). Editing LVLMs faces the challenge of effectively integrating diverse modalities (image and text) while ensuring coherent and contextually relevant modifications. An existing benchmark has three metrics (Reliability, Locality and Generality) to measure knowledge editing for LVLMs. However, the benchmark falls short in the quality of generated images used in evaluation and cannot assess whether models effectively utilize edited knowledge in relation to the associated content. We adopt different data collection methods to construct a new benchmark, $\textbf{KEBench}$, and extend new metric (Portability) for a comprehensive evaluation. Leveraging a multimodal knowledge graph, our image data exhibits clear directionality towards entities. This directional aspect can be further utilized to extract entity-related knowledge and form editing data. We conducted experiments of different editing methods on five LVLMs, and thoroughly analyze how these methods impact the models. The results reveal strengths and deficiencies of these methods and, hopefully, provide insights into potential avenues for future research.