Awesome Test-Time Adaptation

Task

📈 Papers per Year

🧩 Papers per Task (Click to filter)

👤 Top Authors (By published paper count, excl. arXiv; click to filter)

🏷️ Top Tags (Domain & setting; click to filter)

🏛️ Venue × Year (Venues with ≥ 5 papers)

TTA Scope of This Page

✔ Included

Test-time training with self-supervision (TTT)
Entropy minimization / pseudo-labeling at test time (TENT-style)
BN-statistic adaptation, prompt tuning at test time
Online / continual / episodic / single-image TTA
Source-free adaptation done at inference

✘ Excluded

LLM test-time scaling / reasoning with no weight update (best-of-N, CoT)
Source-free DA requiring a separate training phase on the whole target set
Domain generalization (train-time only)
Test-time augmentation only (no model/statistic update)

Two-axis Taxonomy

Every paper has one Task and any number of Tags. The Task is the machine-learning problem it solves. The Tags say which domain it works in and which adaptation setting it uses.

Task (The ML problem, single-select)

Classification	Predict a category: image, VLM, node, graph, audio, tabular, activity, or medical class (also open-set and face anti-spoofing).
Segmentation	Dense per-pixel or per-point labeling.
Detection	Object or region localization.
Regression	Continuous targets such as depth, pose, geometry, time-series forecasting, and molecular properties.
Restoration	Recover a clean signal: super-resolution, denoising, deblurring, dehazing, compression, point-cloud upsampling.
Retrieval & Ranking	Match or rank by similarity: cross-modal retrieval, person re-ID, recommendation.
Sequence & Language	Sequence and language tasks: ASR, OCR, language modeling, QA and reasoning, VQA, machine translation.
Decision-Making	Sequential decision and control: RL, embodied agents, navigation.
Theory	The main contribution is theoretical: learnability, provable guarantees, asymptotics.
Benchmark / Survey	Benchmarks, empirical studies, and surveys.

Tags (Domain & setting, multi-select)

Domain the modality or application area: Image Vision-Language 3D/Point-Cloud Graph Video Speech Audio Text/Document Time-Series Tabular Medical Remote-Sensing EEG/Bio-signal Recommendation Multimodal …

Setting the adaptation protocol and any extra challenge beyond plain distribution shift. Only the distinctive ones are tagged; the plain corruption and source-free case is the unmarked default.

Protocol: Online Continual Single-image Mixed/Non-i.i.d. Federated Black-box Gradient-free Challenge: Open-set Label-Shift Adversarial

Contributing

Spotted a missing paper or an error? Pull requests are welcome. There are two ways to add a paper.

Option A: with Claude Code (recommended)

The repo ships an add-paper skill for Claude Code. Open the repo in Claude Code and just say, for example, "add this paper: <an arXiv link, a conference page, or simply the paper's title>". Given only a title, the skill searches for the paper itself. It then reads the abstract, decides the Task and the Domain/Setting tags, writes the BibTeX, checks for duplicates, appends the entry to papers.json, and regenerates data.json and README.md. You just review the diff and open a PR.

Option B: by hand

Append one object to papers.json with the fields full, title, authors, venue, task, tags, url, pdf, code, bibtex, presentation.
Optionally run python build.py and python readme.py to preview locally.
Open a pull request that commits only papers.json. data.json and README.md are auto-generated and will be regenerated by the maintainer.