The Geometry of Transfer: Unlocking Medical Vision Manifolds for Training-Free Model Ranking
Jiaqi Tang, Shaoyang Zhang, Xiaoqi Wang, Jiaying Zhou, Yang Liu, Qingchao Chen · 2026
Read The Geometry of Transfer: Unlocking Medical Vision Manifolds for Training-Free Model Ranking by Jiaqi Tang, Shaoyang Zhang, Xiaoqi Wang, Jiaying Zhou, Yang Liu, Qingchao Chen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-23916.txt.
Description
Jiaqi Tang and five co-authors present a topology-driven way to judge which medical foundation model will transfer best to a new segmentation task, without fine-tuning each candidate. The 2026 paper argues that existing transferability scores, built for classification on global statistics, miss the topological complexity that dense prediction demands.
The framework scores manifold tractability through three parts. A global divergence uses minimum spanning trees to compare feature and label structure. A local consistency score focuses on anatomical boundaries. A task-adaptive fusion weights global and local signals by the target task's semantic breadth. Tested on the large OpenMind benchmark across varied anatomies and self-supervised models, it beats leading baselines by a large relative margin on rank correlation, offering a cheap, training-free proxy for model selection.
Catalog record
| CREATOR | Jiaqi Tang, Shaoyang Zhang, Xiaoqi Wang, Jiaying Zhou, Yang Liu, Qingchao Chen |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 20 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-23916