Direct Product Flow Matching: Decoupling Radial and Angular Dynamics for Few-Shot Adaptation
Hongxu Chen, Yanghao Wang, Bowei Zhu, Hongxiang Li, Zhen Wang, Ziqi Jiang, Lin Li, Rui Liu, Long Chen · 2026
Read Direct Product Flow Matching: Decoupling Radial and Angular Dynamics for Few-Shot Adaptation by Hongxu Chen, Yanghao Wang, Bowei Zhu, Hongxiang Li, Zhen Wang, Ziqi Jiang, Lin Li, Rui Liu, Long 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-2605-05054.txt.
Description
Hongxu Chen and eight co-authors rethink flow matching methods that adapt vision-language models from only a few examples. Such methods model the alignment between image and text features as a continuous flow. The authors argue existing versions are limited by clashing geometric assumptions about those pre-trained features, which they expose by splitting features into radial and angular parts.
That split reveals three problems: uneven angular speed from coupling, lost confidence information when radial parts are normalised away, and dataset context dropped during feature extraction. Their warped product framework separates radial and angular dynamics on a suitable manifold. The resulting method aims at cleaner few-shot adaptation of vision-language models.
Catalog record
| CREATOR | Hongxu Chen, Yanghao Wang, Bowei Zhu, Hongxiang Li, Zhen Wang, Ziqi Jiang, Lin Li, Rui Liu, Long Chen |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 50 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.05054