MLUBench: A Benchmark for Lifelong Unlearning Evaluation in MLLMs

The article introduces MLUBench, a comprehensive benchmark designed for evaluating lifelong unlearning in multimodal large language models (MLLMs), featuring 127 entities across 9 classes. It highlights the unique challenges of MLLM lifelong unlearning, particularly the preservation of multimodal alignment, and presents LUMoE, a new method that effectively mitigates degradation issues observed in existing unlearning techniques. This benchmark and method are crucial for practitioners addressing data removal requests in MLLMs, ensuring model integrity while adapting to changing data requirements.