Temperature field reconstruction (TFR) is crucial for the thermal management of complex electronic devices. It involves inferring the entire temperature state from sensor observations. Although previous deep learning-based methods perform well in single layout of heat source scenarios, they often overlook the diverse layouts in the ever-changing real world. To fill this gap, this paper proposes a two-stage training framework to reconstruct the temperature field across multiple layouts. First, to acquire effective reconstruction features, we designed a Layouts in-context Encoder (LiE) that operates only on the masked temperature field images, which is essential for feature extraction in the demonstration example. Second, we introduce a feature fusion encoder-decoder architecture. Here, a supervised learning encoder is employed to capture the features of sparse sensor data. By combining the features of the supervised encoder with the layout contextual features of LiE, the decoder can successfully restore the temperature field information. Extensive experiments demonstrate that our method outperforms previous TFR methods in multiple layouts. Furthermore, our approach enables efficient TFR in out-of-distribution (OOD) scenarios.