The authors first probed native models at different Transformer depths.
Native multimodal models can discard most visual tokens with little loss
PACE selects visual tokens by spatial direction and model depth, retaining 25% to 50% while roughly halving reported computation.
Chinese Tech
Lei Tian · Chunzheng Zhu · Kai Wu · Yifan Zhang · Haihua Yang
ByteDance · Hunan University · Institute of Automation, Chinese Academy of Sciences
Research Digest··3 min read
Tian et al.
§