<p>Meta-learning (ML) methods fail to adapt to heterogeneous contexts of unknown tasks during test-time, tend to have meta-overfitting on task-shared structures, and therefore obtain poor generalization performance. In this paper, we propose an algorithm of automatic test-time adaptation for heterogeneous contexts (ATTA-HC) in ML to mitigate meta-overfitting, facilitate minimal adaptation of heterogeneous context, and enhance interpretability. The ATTA-HC algorithm is general, data-agnostic, model-agnostic, and loss-agnostic. Specifically, ATTA-HC automatically achieves this by automatically dividing model parameters after meta-training into two distinct sets based on the quantile of parameters: task-specific parameters describing the heterogeneous contexts of tasks, which are fine-tuned for each individual task, and task-shared parameters, which are meta-trained and encode shared structures across tasks. During test-time, only the task-specific parameters require updating, resulting in a compact task representation. ATTA-HC does not need crafted structures, extra parameters, computation cost, and input manipulation. Experiments including regression, classification, and reinforcement learning show the generalization and superiority of ATTA-HC. Furthermore, our experiments shed light on potential flaws in existing benchmarks, revealing that the level of test-time adaptation can be minimal and automatic in certain scenarios.</p>