Hao Wang

Xinwei Zhang, Hangcheng Liu, Li Bai et al.

Visual token compression is widely used to accelerate large vision-language models (LVLMs) by pruning or merging visual tokens, yet its adversarial robustness remains unexplored. We show that existing encoder-based attacks can substantially overestimate the robustness of compressed LVLMs, due to an optimization-inference mismatch: perturbations are optimized on the full-token representation, while inference is performed through a token-compression bottleneck. To address this gap, we propose the Compression-AliGnEd attack (CAGE), which aligns perturbation optimization with compression inference without assuming access to the deployed compression mechanism or its token budget. CAGE combines (i) expected feature disruption, which concentrates distortion on tokens likely to survive across plausible budgets, and (ii) rank distortion alignment, which actively aligns token distortions with rank scores to promote the retention of highly distorted evidence. Across diverse representative plug-and-play compression mechanisms and datasets, our results show that CAGE consistently achieves lower robust accuracy than the baseline. This work highlights that robustness assessments ignoring compression can be overly optimistic, calling for compression-aware security evaluation and defenses for efficient LVLMs.

Zixuan Xia, Hao Wang, Pengcheng Weng et al.

Multimodal learning aims to integrate complementary information from heterogeneous modalities, yet strong optimization alone does not guaranty well-structured representations. Even under carefully balanced training schemes, multimodal models often exhibit geometric pathologies, including intra-modal representation collapse and sample-level cross-modal inconsistency, which degrade both unimodal robustness and multimodal fusion. We identify representation geometry as a missing control axis in multimodal learning and propose \regName, a lightweight geometry-aware regularization framework. \regName enforces two complementary constraints on intermediate embeddings: an intra-modal dispersive regularization that promotes representation diversity, and an inter-modal anchoring regularization that bounds sample-level cross-modal drift without rigid alignment. The proposed regularizer is plug-and-play, requires no architectural modifications, and is compatible with various training paradigms. Extensive experiments across multiple multimodal benchmarks demonstrate consistent improvements in both multimodal and unimodal performance, showing that explicitly regulating representation geometry effectively mitigates modality trade-offs.