Research map / Label-efficient and robust learning / Domain adaptation and generalization
Exploring Mode Connectivity in Krylov Subspace for Domain Generalization
Aodi Li, Liansheng Zhuang, Xiao Long, Houqiang Li, Shafei 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 |
|---|---|---|---|
| use Krylov-inspired initial direction instead of random initial direction (ablation) | BOA · VLCS LabelMe, loss variation 0.3 | test accuracy: 69.2 -> 70.8; 0.5-1.6% across all loss variation values | Improved |
| add Krylov-subspace-constrained reflection on top of Krylov-inspired p0 (ablation) | BOA · VLCS LabelMe, loss variation 0.3 | test accuracy: 70.8 -> 75.7 | Improved |
| full BOA within Krylov subspace versus random p0 (ablation) | BOA · VLCS LabelMe, loss variation 0.3 | test accuracy: 69.2 -> 75.7 | Improved |
| add BOA on top of SAM with ViT-B/16 prompt tuning | SAM · DomainBed average (ViT-B/16) | OOD accuracy: 76.2 -> 78.1 | Improved |
| apply BOA instead of SAM with ViT-B/32 | ViT-B/32 CLIP VPT · DomainBed average (ViT-B/32) | OOD accuracy: 70.9 -> 73.3 | Improved |
| apply BOA instead of SAM with ViT-L/14 | ViT-L/14 CLIP VPT · DomainBed average (ViT-L/14) | OOD accuracy: 80.1 -> 81.5 | Improved |
| apply BOA instead of SAM with ResNet50 | ResNet50 · VLCS (ResNet50) | OOD accuracy: 80.2 -> 81.8 | Improved |
| use oracle negative test gradient as BOA initial direction (ablation) | ERM and SAM warm-up · DomainBed average (ViT-B/16) | OOD accuracy: 75.5 -> 80.4 (ERM); 76.2 -> 81.0 (SAM) | Improved |
| move along approximate Krylov initial direction (beta_k=1) without reflection (ablation) | SAM warm-up model on VLCS · VLCS domains | test accuracy: +1.4% Caltech101, +6.5% LabelMe, +2.7% SUN09, +2.9% VOC2007 | Improved |
| use Krylov subspace instead of random subspace for approximating test gradient (ablation) | ResNet50 gradient alignment · SUN09 cosine similarity, K=10 | cosine similarity: 0.2657 -> 0.7617 | Improved |
| use Krylov subspace instead of random subspace on Caltech101 (ablation) | ResNet50 gradient alignment · Caltech101 cosine similarity, K=10 | cosine similarity: 0.3125 -> 0.3381 | No clear effect |
| increase Krylov subspace dimension from 1 to 7-10 (ablation) | test gradient alignment on LabelMe (ViT) · LabelMe cosine similarity | cosine similarity: cos < 0.1 at K=1 -> above 0.8 at K=7-10 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CORAL — CORAL: Disentangling Latent Representations in Long-Tailed Diffusion (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Seeking Consistent Flat Minima for Better Domain Generalization via Refining Loss Landscapes (CVPR 2025 · similarity 0.86)
- Generalized Linear Mode Connectivity for Transformers (NeurIPS 2025 · similarity 0.85)
- Topological Generalization Bounds for Discrete-Time Stochastic Optimization Algorithms (NeurIPS 2024 · similarity 0.85)
- Understanding Mode Connectivity via Parameter Space Symmetry (ICML 2025 · similarity 0.85)
- Input Space Mode Connectivity in Deep Neural Networks (ICLR 2025 · similarity 0.85)
- No Wrong Turns: The Simple Geometry Of Neural Networks Optimization Paths (ICML 2024 · similarity 0.84)
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.
