Research map / Label-efficient and robust learning / Test-time adaptation
Test-Time Prompt-Agnostic Decomposition
Junze Wang, Lei Fan, Dezheng Zhang, Donglin Di, Yang Song, Sidong Liu, Cong Cong
Read the paperCheck your idea against 137,217 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 a test-time prompt tuning approach that decomposes prompt updates to be prompt-agnostic, aiming to prevent accumulation of noisy unlabeled objectives across test-time steps.
Problem
Test-time prompt tuning (TPT) adapts large pretrained models to unseen distribution shifts using unlabeled test data; prompt updates are optimized locally at each step, and noisy or biased unlabeled objectives can accumulate across steps and destabilize adaptation.
Approach
Decomposes the prompt update so that it does not depend solely on the current prompt state, reducing error accumulation from reusing updated prompts across test-time steps (based on the abstract only; details unavailable).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TPD on top of visual TPT methods (VPT, E2VPT, VFPT, BPT) | ViT-based visual prompt tuning · Domain generalization OOD mean | accuracy: OOD mean gains 5.51-8.18% | Improved |
| add TPD on top of CLIP text TPT methods (TPT, C-TPT, R-TPT, O-TPT, CAT, DyPt) | CLIP-based text prompt tuning · Domain generalization OOD mean | accuracy: +2.27 to +2.69 | Improved |
| add TPD instead of simple stabilization (clipping, EMA, norm control, low-pass filtering) (ablation) | VPT · DG mean | accuracy: 44.28 -> 47.84 | Improved |
| add TPD instead of low-pass filtering (ablation) | TPT · CSD mean | accuracy: 66.23 -> 68.03 | Improved |
| use diagonal Koopman operator instead of full operator (ablation) | TPD (AKC) · VPT DG mean | accuracy: 43.58 -> 45.92 | Improved |
| use Hankel delay embedding instead of plain projection (ablation) | TPD (HUR) · VPT DG mean | accuracy: 42.31 -> 45.98 | Improved |
| use full routing (persistent 1, oscillatory beta_c, residual beta_n) instead of removing any one routing component (ablation) | TPD HUR routing · VPT and TPT DG/CSD | accuracy: 47.06 -> 47.84 (VPT DG, w/o Pres.) | Improved |
| add TPD to prompt-tuning methods for calibration | TPT, C-TPT and others · ImageNet-A ECE | ECE: 16.42% -> 12.27% | Improved |
| add TPD to TPT on ImageNet-R for calibration | TPT · ImageNet-R ECE | ECE: 4.36% -> 1.62% | Improved |
| add TPD to VPT across batch sizes | VPT · ImageNet-A batch-size sweep | accuracy: 1.14% -> 12.82% at B=1; decrease at B=8 | Mixed |
| increase batch size to 128 with TPD (ablation) | VPT + TPD · ImageNet-A | accuracy: 19.17% -> 38.98% at B=128 | Improved |
| naive entropy-minimization test-time tuning of visual prompts without stabilization | VPT · Domain generalization OOD mean | accuracy: 46.17 -> 39.66 | Hurt |
Experimental setup
Builds on: test-time prompt tuning
Closest papers by meaning
- DynaPrompt: Dynamic Test-Time Prompt Tuning (ICLR 2025 · similarity 0.88)
- Black-Box Test-Time Prompt Tuning for Vision-Language Models (AAAI 2025 · similarity 0.87)
- DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.87)
- Robust Test-Time Adaptation for Zero-Shot Prompt Tuning (AAAI 2024 · similarity 0.86)
- Revisiting the Power of Prompt for Visual Tuning (ICML 2024 · similarity 0.86)
- Revisiting Prefix-tuning: Statistical Benefits of Reparameterization among Prompts (ICLR 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.
