[arXiv]score: 0.13

MLUBench Reveals Cumulative Degradation in Multimodal Lifelong Unlearning

June 12, 2026

MLUBench introduces a 127-entity, 9-class benchmark for evaluating sequential data removal in multimodal LLMs, finding that existing unlearning methods suffer severe cumulative degradation across sequential requests. A key finding is that multimodal alignment must be preserved during unlearning, a constraint absent in unimodal settings that current methods fail to handle.

HOW THIS AFFECTS YOU

●

researcherThe multimodal alignment preservation constraint is a concrete new failure mode for unlearning research that unimodal benchmarks cannot surface.

●

policySequential data removal requests — as required under GDPR and similar regulations — are shown to be practically unsolved for multimodal models, which has direct compliance implications.

read original ↗arxiv.org

← back to feed