Workshop: Scale to See: Scalable AI for Scientific Imaging and Spatiotemporal Data
Authors: Ye Tian (The University of Hong Kong), Yiming Yin and Xiaowen Chu (Hong Kong University of Science and Technology (Guangzhou)), Yizhou Yu (The University of Hong Kong), Hong Xu (The Chinese University of Hong Kong), Bo Li (Hong Kong University of Science and Technology), and Chuan Wu (The University of Hong Kong)
Abstract: High-resolution chest radiographs produce long visual-token sequences that increase large vision-language model (LVLM) fine-tuning cost and inference prefill latency. We present \textbf{A}nchor-\textbf{C}ontext \textbf{S}implified \textbf{T}oken \textbf{R}etention and \textbf{R}eduction (AC-STRR), a training-aware framework that filters background tokens, preserves locally distinctive anchors, and merges redundant context before the language decoder. AC-STRR applies the same reduction path during fine-tuning and inference while retaining the original position indices used by Multimodal Rotary Position Embedding. On VinDr-CXR with Qwen3-VL-8B and Low-Rank Adaptation, AC-STRR reduces fine-tuning time on one L40S by 24.9\% and achieves a 9.07$\times$ combined speedup on eight L40S GPUs over the single-GPU full-token baseline. It also reaches up to 63.5\% token reduction on native-resolution MIMIC-CXR. Tuned AC-STRR keeps report quality close to the full-token baseline while lowering prefill latency. These results suggest that domain-aware token reduction can improve fine-tuning and inference prefill efficiency for high-resolution medical LVLMs.
Back to Scale to See: Scalable AI for Scientific Imaging and Spatiotemporal Data Archive Listing
Back to Full Workshop Archive Listing