Visual Token Compression Enhances Robustness of MLLMs

First demonstration that visual token pruning reduces vulnerabilities in multimodal LLMs, including jailbreak attacks and hallucinations. The method compresses visual tokens while preserving key information, improving model safety.