Test-time adaptation (TTA) leverages testing data to enhance model intelligence, demonstrating significant potential for advancing models capabilities from system-1/1.5 intelligence to higher. However, conventional TTA typically relies on model updates via gradient back-propagation. This approach introduces challenges including instability, inflexibility, and computational burdens—limitations that become increasingly problematic as models scale and are often unsuitable for resource-constrained scenarios.

In this talk, we will first explore how relaxing core assumptions of TTA enables robust performance under diverse test conditions, such as mixed domain shifts, single-sample adaptation, and online label shifts. Second, we will discuss designing a versatile learning framework applicable to both classification and regression tasks across image-, object-, and pixel-level predictions. Finally, we will present methods to achieve the benefits of learning-based TTA using only forward propagation, making adaptation feasible for quantized models and deployment on resource-limited devices.