Research map / Label-efficient and robust learning / Test-time adaptation
Tail Cues, Principal Corrections: Plug-and-Play Rectification for Open-Set Test-Time Adaptation
Yingkai Yang, Chaoqi Chen, Hui Huang
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 TALOR, a lightweight backward-free plug-and-play rectification module for open-set test-time adaptation that uses the source linear head as a geometric anchor.
Problem
Open-set test-time adaptation on non-stationary streams where covariate and semantic shifts coexist; existing TTA methods update parameters or maintain mutable memories, which is costly and vulnerable to unknown samples.
Approach
A backward-free rectification module that uses the source linear head as a geometric anchor to correct predictions, avoiding parameter updates and mutable memory banks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TALOR as a plug-in on the frozen source model | Source ViT-B/16 · ImageNet-C open-set (mean H-score) | H-score: 56.4 -> 66.1 | Improved |
| add TALOR post-correction to COME | COME · ImageNet-C open-set (mean H-score) | H-score: 63.7 -> 68.4 | Improved |
| add TALOR to SAR | SAR · ImageNet-C open-set (mean H-score) | H-score: 51.1 -> 65.4 | Improved |
| add TALOR to EATA | EATA · Textures-C AUC | AUC: -1.7 AUC on Textures-C; +5.2 mean H-score | Mixed |
| add TALOR to base methods on LAION-C | CoTTA, EATA, OSTTA, ViDA · LAION-C mean H-score | H-score: +10.8, +7.4, +11.5, +10.7 | Improved |
| add TALOR in closed-set protocols | strong baselines such as DPCore · closed-set LAION-C accuracy | accuracy: 65.3 (+0.1) | Improved |
| use only the principal mean term (remove tail slope) (ablation) | TALOR on Source · ImageNet-C open-set mean | H-score: 66.1 -> 62.3 | Hurt |
| use only the tail-to-principal slope term (remove mean) (ablation) | TALOR on Source · ImageNet-C open-set mean | H-score: 66.1 -> 61.4 | Hurt |
| add soft ID/OOD routing instead of one affine rectifier (ablation) | TALOR on Source · ImageNet-C open-set mean | H-score: 65.4 -> 66.1 | Improved |
| use ground-truth routing instead of unsupervised GMM routing (ablation) | TALOR on Source · ImageNet-C open-set mean | H-score: 66.1 -> 66.9 | Improved |
| vary SVD energy threshold tau (ablation) | TALOR · ImageNet-C open-set | H-score | No clear effect |
| remove EMA smoothing (very small beta) (ablation) | TALOR · ImageNet-C open-set | H-score | Hurt |
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.
- STAMP — STAMP: Outlier-Aware Test-Time Adaptation with Stable Memory Replay (ECCV 2024) · 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
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.88)
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.87)
- Self-Bootstrapping for Versatile Test-Time Adaptation (ICML 2025 · similarity 0.86)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.86)
- Free on the Fly: Enhancing Flexibility in Test-Time Adaptation with Online EM (CVPR 2025 · similarity 0.86)
- A Layer Selection Approach to Test Time Adaptation (AAAI 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.
