Research map / Label-efficient and robust learning / Domain adaptation and generalization
DA-MergeLoRA: Hypernetwork-Based LoRA Merging for Few-Shot Test-Time Domain Adaptation
Siobhan Reid, Zhixiang Chi, Li Gu, Omid Heidari, Ziqiang Wang, Yang Wang
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 |
|---|---|---|---|
| merge per-domain LoRAs with a meta-learned hypernetwork instead of using prior FSTT-DA method L2C | CLIP ViT-B/16 · DomainNet average | accuracy: 61.2 -> 62.44 | Improved |
| use hypernetwork merging instead of L2C on FMoW worst-case | CLIP ViT-B/16 · FMoW OOD | WC accuracy: 40.9 -> 52.26 | Improved |
| use hypernetwork merging instead of L2C on iWildCam F1 | CLIP ViT-B/16 · iWildCam OOD | macro F1: 35.2 -> 33.62 | Hurt |
| use hypernetwork merging instead of per-domain best on Quickdraw | CLIP ViT-B/16 · DomainNet Quickdraw | accuracy: 18.7 -> 17.87 | Hurt |
| use learned hypernetwork merge weights instead of simple parameter averaging (ablation) | source LoRAs · FMoW OOD | WC accuracy: 20.63 -> 52.26 | Improved |
| use hypernetwork instead of few-shot AdaMerging of LoRAs (ablation) | source LoRAs · iWildCam OOD | macro F1: 18.66 -> 33.62 | Improved |
| use hypernetwork merging instead of simple averaging on a saturated dataset (ablation) | source LoRAs · Camelyon17 OOD | accuracy: 95.26 -> 96.90 | Improved |
| use per-domain LoRAs with merging instead of one universal LoRA (ablation) | CLIP ViT-B/16 · FMoW OOD | WC accuracy: 51.02 -> 52.26 | Improved |
| condition hypernetwork on target images instead of noise (ablation) | hypernetwork · iWildCam OOD | macro F1: 17.39 -> 33.62 | Improved |
| add domain tokens to LoRA columns (ablation) | hypernetwork · iWildCam OOD | macro F1: 25.72 -> 33.62 | Improved |
| add layer and layer-type tokens (ablation) | hypernetwork · FMoW OOD | WC accuracy: 49.16 -> 52.26 | Improved |
| remove architectural components (tokens, target images) on Camelyon17 (ablation) | hypernetwork · Camelyon17 OOD | accuracy: 96.92 vs 96.90 | No clear effect |
Closest papers by meaning
- IterIS: Iterative Inference-Solving Alignment for LoRA Merging (CVPR 2025 · similarity 0.88)
- LoRA-X: Bridging Foundation Models with Training-Free Cross-Model Adaptation (ICLR 2025 · similarity 0.88)
- Test-Time Domain Adaptation by Learning Domain-Aware Batch Normalization (AAAI 2024 · similarity 0.88)
- LoRA Recycle: Unlocking Tuning-Free Few-Shot Adaptability in Visual Foundation Models by Recycling Pre-Tuned LoRAs (CVPR 2025 · similarity 0.87)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.87)
- Learning to Adapt Frozen CLIP for Few-Shot Test-Time Domain Adaptation (ICLR 2025 · similarity 0.87)
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.
