6 papers
DecoMoE: Decoupling Visual Propagation and Expert Computation for Efficient Multimodal MoE Inference
Xudong Tan, Peng Ye, Ming Xie +4
Multimodal mixture-of-experts (MoE) models combine sparse expert activation with visual-language capabilities, yet their inference remains costly because long visual-token sequence…
Revisiting Multimodal KV Cache Compression: A Frequency-Domain-Guided Outlier-KV-Aware Approach
Yaoxin Yang, Peng Ye, Xudong Tan +4
Multimodal large language models suffer from substantial inference overhead since multimodal KV Cache grows proportionally with the visual input length. Existing multimodal KV Cach…
Sequential Token Merging: Revisiting Hidden States
Yan Wen, Peng Ye, Lin Zhang +4
Vision Mambas (ViMs) achieve remarkable success with sub-quadratic complexity, but their efficiency remains constrained by quadratic token scaling with image resolution. While exis…
Think Twice, Act Once: Token-Aware Compression and Action Reuse for Efficient Inference in Vision-Language-Action Models
Xudong Tan, Yaoxin Yang, Peng Ye +5
Vision-Language-Action (VLA) models have emerged as a powerful paradigm for general-purpose robot control through natural language instructions. However, their high inference cost-…
TokenCarve: Information-Preserving Visual Token Compression in Multimodal Large Language Models
Xudong Tan, Peng Ye, Chongjun Tu +5
Multimodal Large Language Models (MLLMs) are becoming increasingly popular, while the high computational cost associated with multimodal data input, particularly from visual tokens…
Multi-Level Decoupled Relational Distillation for Heterogeneous Architectures
Yaoxin Yang, Peng Ye, Weihao Lin +4
Heterogeneous distillation is an effective way to transfer knowledge from cross-architecture teacher models to student models. However, existing heterogeneous distillation methods…