Test-Time Training (TTT) improves model robustness under distribution shifts by adapting the model at inference time using unlabeled test data, enabling better generalization to dynamic or unseen environments. Most TTT methods apply self-supervised tasks such as entropy minimization or contrastive learning to features from a single layer, often overlooking the contribution of features from other layers to model robustness. In this work, we argue that both shallow and deep features play essential roles in achieving robust representations for TTT. Motivated by this observation, we propose a novel multi-layer noise contrastive test-time training framework, which applies noise contrastive learning to all intermediate layers of a deep neural network. Specifically, we inject Gaussian noise with different variances into the outputs of each layer in a ResNet backbone, and employ dedicated discriminators to distinguish the noise levels, thereby enforcing robustness across the entire feature hierarchy. Extensive experiments on domain adaptation and corruption benchmarks show our method consistently outperforms existing TTT approaches, achieving state-of-the-art results and demonstrating the effectiveness of multi-layer noise contrastive learning for robust TTT.