MLUBench: A Benchmark for Evaluating Continual Unlearning in Multimodal Large Language Models

Researchers propose MLUBench, a large-scale benchmark for continual unlearning in multimodal large language models (MLLMs), covering 127 entities across 9 categories, with code and dataset open-sourced at https://github.com/lihe-maxsize/Lifelong_Unlearning_main。针对随时间依次到达的数据删除请求，实验显示现有遗忘方法出现严重且累积的退化。论文指出该问题的独特挑战：与单模态模型不同，MLLM. Continual unlearning is constrained by the preservation of multimodal alignment; continually unlearning from a single modality may harm the entire model. To this end, the authors propose the LUMoE method, and experiments show that it significantly mitigates the degradation faced by baseline methods.