[论文] CellTrack-Bench: Microscopic Cell Tracking Benchmark for Vision-Langua...

论文概要

研究领域: CV 作者: Ilan Theodoro, Jordão Bragantini, Loïc A. Royer 发布时间: 2026-10-22 arXiv: 2610.31566

中文摘要

大语言模型在文本优先任务上表现出色，但在高度专业化科学领域的通用准确性仍存疑。细胞追踪需要对细胞动态的时空感知。CellTrack-Bench评估视觉语言模型与专门追踪器相比的细胞追踪性能：来自15个标准数据集、四种成像模态，包含1,537个问题和9,041个细胞连接决策。通过识别不同算法追踪器共有的失败选择困难子集。评估了多种环境工具，基准测试9个前沿闭源模型和13个开放权重模型，并提供了详细的失败模式分析。

原文摘要

Large language models have shown outstanding performance on text-first tasks, but their general-purpose accuracy in specialized scientific domains remains in question. Cell tracking requires spatiotemporal awareness of cell dynamics. We created CellTrack-Bench, which assesses VLMs' cell-tracking performance against specialized trackers. The benchmark draws from 15 standard cell-tracking datasets across four imaging modalities, comprising 1,537 problems and 9,041 cell-linking decisions. A hard subset was selected by identifying failures shared across distinct algorithmic trackers. We evaluated different environment harnesses, benchmarked 9 frontier closed-source and 13 open-weight models, and provided detailed failure mode analysis.

#论文 #arXiv #CV #小凯