Visual Token Compression Enhances Robustness of MLLMs
The paper argues that pruning visual tokens in multimodal LLMs can improve robustness, because tokens that are poorly aligned with the language feature space behave like out-of-distribution inputs and can contribute to jailbreak susceptibility and hallucinations. Its neat angle is treating visual-token compression not just as an efficiency trick but as a safety mechanism, with a simple distance-based iterative pruning scheme that improves jailbreak defense and stays competitive on hallucination/general benchmarks, though the abstract leaves open how reliable that distance heuristic is across architectures and whether pruning can drop subtle but task-relevant visual evidence.
Abstract
In this paper, we show for the first time that visual token pruning enhances the robustness of Multimodal Large Language Models (MLLMs), mitigating vulnerabilities such as jailbreak attacks and hallucinations. Given that vision and language modalities cannot be perfectly aligned, the misaligned visual tokens might act as out-of-distribution (OOD) inputs, leading to unpredictable outputs and introducing potential vulnerabilities. Building on this insight, we aim to enhance model robustness against jailbreaks and hallucinations by reducing OOD visual tokens at robust-pruning layers, while also reducing inference cost as a side benefit. Specifically, we measure the distance between each visual token and the language feature space. Then, visual tokens with large distances are identified as OOD tokens, which can be iteratively pruned. To demonstrate the effectiveness of our method, we evaluate it on seven diverse popular benchmarks. Notably, our method yields an average improvement of 13.29\% in defending jailbreak attacks, consistently achieves competitive performance in mitigating hallucinations, and maintains strong results on general datasets like MME.