TestDG: Test-time Domain Generalization for Continual Test-time Adaptation

September 14, 2026 | Source: arXiv AI

Tags: test-time-adaptation, domain-generalization, distribution-shift, continual-learning, TestDG

TestDG achieves state-of-the-art on four continual test-time adaptation benchmarks by learning domain-invariant features at inference time — without any access to source data — and shows superior generalization to domains never seen during either training or testing.

Details

Continual test-time adaptation (CTTA) addresses one of the hardest real-world ML deployment problems: models encounter a continuous stream of shifting domains without any ability to retrain. Most CTTA methods adapt to the current domain but forget what they learned about previous ones. TestDG takes a different angle. Instead of optimizing for the current domain, it learns features that are invariant across both the current and past test domains encountered so far. The intuition: features stable across all seen domains are more likely to generalize to unseen future ones. The framework introduces a new model architecture and test-time adaptation strategy specifically designed for domain-invariant feature learning, along with prototype selection and update mechanisms for managing accumulated domain information efficiently. Results across four public CTTA benchmarks show state-of-the-art performance. The paper also evaluates generalization to held-out domains never seen during evaluation — the more practically relevant test. This is v3 of the paper (originally submitted April 2025), with updated content.