Research map / Architectures and efficiency / Pruning and efficient tuning
DaWin: Training-free Dynamic Weight Interpolation for Robust Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DaWin, a training-free dynamic weight interpolation method that sets per-sample interpolation coefficients from the entropy ratio of individual models and uses a Beta/Dirichlet mixture model to reduce inference cost.
Problem
Static weight interpolation (e.g., WiSE-FT, model merging) uses one global coefficient for all test samples, which limits downstream accuracy under distribution shift and in multi-task merging; existing dynamic alternatives need extra training or routers.
Approach
Estimate each model's per-sample expertise by exponentiated negative prediction entropy on unlabeled test inputs, convert it to a per-sample interpolation coefficient, and cluster coefficients with a Beta mixture (or Dirichlet for multi-task) so only K merged models are needed at inference; no training is required.
Key results
- On multi-task learning with CLIP ViT-B/32 over 8 tasks, DaWin reaches 86.3% average accuracy, versus 84.8% for Pareto Merging and 80.0% for AdaMerging.
- On ImageNet with CLIP ViT-B/32, DaWin (w/o mixture modeling) gets 54.41% average OOD accuracy versus 52.08% for Uniform Soup and 51.02% for WiSE-FT, with ID accuracy 78.70%.
- Oracle sample-wise dynamic interpolation reaches 60.6% average over ImageNet shifts versus 51.0% for WiSE-FT, motivating the approach.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace static global interpolation with sample-wise oracle interpolation | WiSE-FT · ImageNet distribution shifts | accuracy: 51.0 -> 60.6 | Helped |
| use sample-wise entropy-based dynamic interpolation instead of static merging | CLIP ViT-B/32 robust fine-tuning · OOD average | accuracy: 51.02 -> 54.36 | Helped |
| use sample-wise entropy-based dynamic interpolation instead of Pareto Merging | CLIP ViT-B/32 multi-task learning · Multi-task average | accuracy: 84.8 -> 86.3 | Helped |
| remove mixture modeling (ablation) | DaWin · OOD average | accuracy: 54.36 -> 54.41 | No clear effect |
| add domain-wise offset adjustment (ablation) | DaWin · ImageNet ID | accuracy: 78.3 -> 78.7; 83.1 -> 83.4; 86.7 -> 86.9 | Helped |
| add domain-wise offset adjustment (ablation) | DaWin · OOD average | accuracy: 54.1 -> 54.4; 66.6 -> 66.9; 75.9 -> 76.0 | Helped |
| replace entropy with pseudo-label expertise metrics (ablation) | DaWin · ImageNet ID and OOD | accuracy | Mixed |
| vary the sample size used to fit the Beta mixture model (ablation) | DaWin on ImageNet and distribution shifts · ImageNet and OOD average | accuracy: 78.55 / 54.30; 78.69 / 54.27; 78.71 / 54.27; 78.70 / 54.28; 78.70 / 54.25; 78.71 / 54.24; 78.67 / 54.26; 78.70 / 54.36 | No clear effect |
| remove mixture modeling (ablation) | DaWin · ImageNet ID | accuracy: 78.70 -> 78.71 | No clear effect |
Surprising findings
- Per-sample entropy-based coefficients nearly match oracle per-sample X-entropy coefficients on most shifts, but diverge on ImageNet-A where entropy correlates only weakly with X-entropy.
- Adding WiSE-FT interpolation on top of DaWin gives little gain for ViT-B/16 and ViT-L/14, suggesting DaWin already lies beyond the Pareto frontier of the two models.
Limitations
- Relies on entropy as a proxy for cross-entropy, which is weak when both models fail (FalseFalse split) or on ImageNet-A.
- Inference cost still grows with the number of models and clusters (O(K+M)) and requires batch-level test data for the mixture fit.
- Offset adjustment assumes access to the whole test set per domain, which is not always available in streaming settings.
Experimental setup
Datasets: ImageNet-1K ImageNet-V2 ImageNet-R ImageNet-A ImageNet-Sketch ObjectNet SUN397 Cars RESISC45 EuroSAT SVHN GTSRB
Benchmarks: ImageNet-1K ImageNet-V2 ImageNet-R ImageNet-A ImageNet-Sketch ObjectNet SUN397 Cars RESISC45 EuroSAT SVHN GTSRB
Metrics: top-1 accuracy average OOD accuracy wall-clock inference time mean squared error of coefficients
Baselines: ZS FT WiSE-FT Output ensemble Uniform Soup Greedy Soup Model Stock LP-FT CAR-FT FLYP Lipsum-FT CaRot
Builds on: WiSE-FT Task Arithmetic AdaMerging Model Soups dynamic classifier selection
Closest papers by meaning
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.85)
- Dynamic Model Merging Made Slim (NeurIPS 2026 · similarity 0.85)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.84)
- Time- Memory- and Parameter-Efficient Visual Adaptation (CVPR 2024 · similarity 0.84)
- ADDP: Learning General Representations for Image Recognition and Generation with Alternating Denoising Diffusion Process (ICLR 2024 · similarity 0.84)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · 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.