Anzhou Hou

Zi-Hao Bo, Yaqian Li, Anzhou Hou et al.

Mixture-of-Experts (MoE) has become a prevalent backbone for large vision-language models (VLMs), yet how modality-specific signals should guide expert routing remains under-explored. Existing routing strategies are either hand-crafted or modality-agnostic, relying on idealized priors that ignore the layer-dependent modality fusion patterns in MoE-VLMs and provide little guidance for expert specialization. We propose Soft Modality-guided Expert Specialization (SMoES), which consists of dynamic soft modality scores that capture layer-dependent fusion patterns, an expert binning mechanism aligned with expert-parallel deployment, and an inter-bin mutual information regularization that encourages coherent modality specialization. Our method leverages attention-based or Gaussian-statistics modality scores to optimize mutual information regularization. Experiments across four MoE-based VLMs and 16 benchmarks demonstrate improvement on both effectiveness and efficiency: 0.9% and 4.2% average gain on multimodal and language-only tasks, 56.1% reduction in EP communication overhead, and 12.3% throughput improvement under realistic deployment. These results validate that aligning routing with modality-aware expert specialization unlocks MoE-VLM capacity and efficiency.

Rinyoichi Takezoe, Yaqian Li, Zihao Bo et al.

Vision-Language Models (VLMs) have recently demonstrated remarkable capabilities in visual understanding and reasoning, but they also impose significant computational burdens due to long visual sequence inputs. Recent works address this issue by pruning unimportant visual tokens, achieving substantial computational reduction while maintaining model performance. The core of token pruning lies in determining token importance, with current approaches primarily relying on attention scores from vision encoders or Large Language Models (LLMs). In this paper, we analyze the effectiveness of attention mechanisms in both vision encoders and LLMs. We find that vision encoders suffer from attention sink, leading to poor focus on informative foreground regions, while in LLMs, although prior studies have identified attention bias toward token positions, text-to-vision attention demonstrates resistance to this bias and enables effective pruning guidance in middle layers. Based on these observations, we propose LearnPruner, a two-stage token pruning framework that first removes redundant vision tokens via a learnable pruning module after the vision encoder, then retains only task-relevant tokens in the LLM's middle layer. Experimental results show that our LearnPruner can preserve approximately 95% of the original performance while using only 5.5% of vision tokens, and achieve 3.2$\times$ inference acceleration, demonstrating a superior accuracy-efficiency trade-off.

Jinyue Bian, Zhaoxing Zhang, Zhengyu Liang et al.

The Visual-Language-Action (VLA) models can follow text instructions according to visual observations of the surrounding environment. This ability to map multimodal inputs to actions is derived from the training of the VLA model on extensive standard demonstrations. These visual observations captured by third-personal global and in-wrist local cameras are inevitably varied in number and perspective across different environments, resulting in significant differences in the visual features. This perspective heterogeneity constrains the generality of VLA models. In light of this, we first propose the lightweight module VLA-LPAF to foster the perspective adaptivity of VLA models using only 2D data. VLA-LPAF is finetuned using images from a single view and fuses other multiview observations in the latent space, which effectively and efficiently bridge the gap caused by perspective inconsistency. We instantiate our VLA-LPAF framework with the VLA model RoboFlamingo to construct RoboFlamingo-LPAF. Experiments show that RoboFlamingo-LPAF averagely achieves around 8% task success rate improvement on CALVIN, 15% on LIBERO, and 30% on a customized simulation benchmark. We also demonstrate the developed viewadaptive characteristics of the proposed RoboFlamingo-LPAF through real-world tasks.