Preprint
Aug 2026
OmniPack: Unified Token Compression for Efficient Omni-modal Large Language Models
OmniPack is proposed, a training-free framework that coordinates structural compression before the LLM with task-relevant semantic refinement within the LLM, and consistently achieves the best performance-efficiency trade-off across diverse retention ratios, outperforming all existing methods.
Wan-Shun Su, Yang Shi, Fei Liu et al.
· 2 citations