MLUBench: A Benchmark for Lifelong Unlearning Evaluation in MLLMs

In three linesMLUBench is a large-scale benchmark for evaluating lifelong unlearning in multimodal large language models (MLLMs). It contains 127 entities across 9 classes. Authors show existing unlearning methods suffer cumulative degradation and propose LUMoE to preserve multimodal alignment during sequential forgetting.

Summary generated by Claude — human-verified