Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Path-Decoupled Hyperbolic Flow Matching for Few-Shot Adaptation
Lin Li, Ziqi Jiang, Gefan Ye, Zhenqi He, Jiahui Li, Jun Xiao, Kwang-Ting Cheng, Long Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add centripetal hyperbolic alignment (CHA) to CLIP-LoRA (ablation) | CLIP-LoRA · difficult group avg, 16-shot | top-1 accuracy: 76.1 -> 77.7 | Improved |
| add path-decoupled objective to flow matching (ablation) | HFM (CHA) · difficult group avg, 16-shot | top-1 accuracy: 77.7 -> 79.2 | Improved |
| add diameter-based stopping at inference (ablation) | HFM (CHA+PO) · difficult group avg, 16-shot | top-1 accuracy: 79.2 -> 79.8 | Improved |
| add entailment loss to the contrastive loss in CHA (ablation) | HFM CHA · difficult group avg, 16-shot | top-1 accuracy: 72.0 -> 79.8 | Improved |
| add contrastive loss to the entailment loss in CHA (ablation) | HFM CHA · difficult group avg, 16-shot | top-1 accuracy: 79.0 -> 79.8 | Improved |
| add inter-class decoupling loss to step-wise consistency loss (ablation) | HFM PO · difficult group avg, 16-shot | top-1 accuracy: 79.4 -> 79.8 | Improved |
| use hyperbolic geometry instead of Euclidean with same objectives (FMA*) (ablation) | flow matching adaptation · difficult group avg, 16-shot | top-1 accuracy: 77.9 -> 79.8 | Improved |
| add HFM objectives to Euclidean FMA (ablation) | FMA · difficult group avg, 16-shot | top-1 accuracy: 77.7 -> 77.9 | No clear effect |
| step-wise consistency supervision instead of velocity supervision in Euclidean FMA (ablation) | FMA · difficult group avg, 16-shot | top-1 accuracy: 77.7 -> 77.8 | No clear effect |
| step-wise consistency supervision instead of velocity supervision in HFM (ablation) | HFM · difficult group avg, 16-shot | top-1 accuracy: 78.8 -> 79.8 | Improved |
| plug HFM into CLIP-Adapter | CLIP-Adapter · difficult group avg, 16-shot | top-1 accuracy: 63.6 -> 72.3 | Improved |
| plug HFM into CLIP-LoRA with ViT-L/14 | CLIP-LoRA · difficult group avg, 16-shot, ViT-L/14 | top-1 accuracy: 81.0 -> 83.1 | Improved |
Closest papers by meaning
- Direct Product Flow Matching: Decoupling Radial and Angular Dynamics for Few-Shot Adaptation (NeurIPS 2026 · similarity 0.90)
- Geometry-Aware Image Flow Matching (ICML 2026 · similarity 0.90)
- Metric Flow Matching for Smooth Interpolations on the Data Manifold (NeurIPS 2024 · similarity 0.87)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.86)
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.86)
- Flow Matching on General Geometries (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
