Research map / Label-efficient and robust learning / Test-time adaptation
FOZO: Forward-Only Zeroth-Order Prompt Optimization for Test-Time Adaptation
Xingyu Wang, Tao 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 |
|---|---|---|---|
| add deep-shallow feature statistics alignment loss (ablation) | FOZO · ImageNet-C (5k, level 5) component ablation | Top-1 accuracy (%): 57.3 -> 60.1 | Improved |
| add dynamic perturbation scale decay (ablation) | FOZO · ImageNet-C (5k, level 5) component ablation | Top-1 accuracy (%): 60.1 -> 62.7 | Improved |
| add zeroth-order prompt optimization with entropy loss (ablation) | ViT-Base · ImageNet-C (5k, level 5) component ablation | Top-1 accuracy (%): 55.1 -> 57.3 | Improved |
| dynamic perturbation speeds convergence (ablation) | FOZO · ImageNet-C level 5 convergence | accuracy vs time | Improved |
| FOZO vs FOA (CMA-ES prompts) at 2 forward passes | forward-only TTA · ImageNet-C (5k, level 5, continual), 2 FP | Top-1 accuracy (%): 58.13 -> 59.52 | Improved |
| FOZO vs ZOA at 2 forward passes | forward-only TTA · ImageNet-C (5k, level 5, continual), 2 FP | Top-1 accuracy (%): 58.56 -> 59.52 | Improved |
| FOZO vs backprop methods (EATA) at 28 forward passes | TTA · ImageNet-C (5k, level 5, continual), 28 FP vs backprop | Top-1 accuracy (%): 61.35 -> 62.67 | Improved |
| FOZO on quantized INT8 model vs FOA | quantized ViT-Base · ImageNet-C (5k, level 5), INT8 quantized ViT-Base | Top-1 accuracy (%): 57.07 -> 58.00 | Improved |
| increase prompts from 3 to 8 (ablation) | FOZO · ImageNet-C Gaussian noise level 5, hyperparameters | Top-1 accuracy (%) | No clear effect |
| increase batch size from 32 to 128 (ablation) | FOZO · ImageNet-C Gaussian noise level 5, hyperparameters | Top-1 accuracy (%): 62.3 -> 61.2 | Hurt |
| reduce batch size to 4 (ablation) | FOZO · ImageNet-C Gaussian noise level 5, hyperparameters | Top-1 accuracy (%): 62.3 -> 60.9 | Hurt |
Later papers that compared with this method
1 later accepted paper reports FOZO as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- EVA-0: Test-Time Model Evolution with Only Two Forward Passes per Sample (NeurIPS 2026 · better on 1 of 1 shared result)
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.
- TENT — Tent: Fully Test-Time Adaptation by Entropy Minimization (ICLR 2021) · 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
- Test-Time Model Adaptation with Only Forward Passes (ICML 2024 · similarity 0.89)
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.88)
- EVA-0: Test-Time Model Evolution with Only Two Forward Passes per Sample (NeurIPS 2026 · similarity 0.87)
- L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer (NeurIPS 2024 · similarity 0.86)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.86)
- Refining Adaptive Zeroth-Order Optimization at Ease (ICML 2025 · 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.
