DOME: Learning Transferable Domain Variables from Sparse Supervision for Test-Time Adaptation

Researchers introduce DOME, a domain encoder that improves test-time adaptation by explicitly modeling sample-specific domain shifts rather than inferring a single global distribution. The method leverages vision-language pretraining and sparse domain banks to achieve state-of-the-art performance on multiple benchmarks, suggesting that structured domain representation outweighs algorithmic complexity.

DOME addresses a fundamental limitation in current test-time adaptation approaches: the assumption that domain shifts are monolithic rather than multidimensional and sample-specific. Traditional TTA methods treat domains as single global distributions, which fails to capture the granular variations in real-world data. The researchers' insight that effective adaptation depends more on explicit domain representation than algorithmic sophistication challenges prevailing assumptions in the field.

This work builds on decades of domain adaptation research but represents a paradigm shift toward zero-shot domain encoding. By leveraging pretrained vision-language models, DOME extracts dense, continuous representations that capture nuanced domain characteristics. The momentum-updated sparse domain bank acts as a structured memory system, enabling disentangled supervision without requiring labeled adaptation data—a critical advantage for practical deployment.

For the AI research community, DOME's superior performance across ImageNet-C, ImageNet-R, and ImageNet-Sketch suggests that practitioners may be overcomplicating adaptation strategies. The finding that basic entropy-minimization paired with explicit domain cues outperforms complex methods has significant implications for model robustness in production environments. This could accelerate adoption of simpler, more interpretable adaptation pipelines in computer vision systems.

Looking ahead, the methodology raises questions about how domain representation principles could transfer to other modalities and tasks. The sparse domain bank concept may inspire new approaches to continual learning and out-of-distribution generalization. Researchers should investigate whether similar explicit representation strategies could improve performance in other adaptation scenarios, potentially reshaping how practitioners approach model robustness.

• →DOME models domain shifts as sample-specific variables rather than global distributions, improving test-time adaptation performance.
• →Vision-language pretraining combined with sparse domain banks enables zero-shot domain encoding without labeled adaptation data.
• →State-of-the-art results achieved using basic entropy minimization, suggesting explicit domain representation matters more than algorithmic complexity.
• →The approach achieves strong performance across multiple benchmarks including ImageNet-C, ImageNet-R, and ImageNet-Sketch.
• →Findings suggest practitioners may be overengineering TTA methods and could benefit from structured domain representation frameworks.