In this work, we present our solution and experiment results for the Multi-Task Learning (MTL) Challenge of the 7th Affective Behavior Analysis in-the-wild (ABAW7) Competition. This challenge consists of three tasks: Action Unit (AU) detection, Facial Expression (EXPR) recognition, and Valance-Arousal (VA) estimation. We address above tasks from three aspects: 1) To learn robust facial feature representations, we first exploit the pre-trained large model DINOv2 to encode rich facial expressions; 2) We design a task-adaptive module (TAM) to learn the discriminative feature representations for each task in a self-adaptive manner. More specifically, we construct a set of learnable query vectors to capture task-specific representation via cross-attention learning; 3) We propose the AU-assisted Graph (AUG) module to capture the inherent correlation between AUs and then apply it to assist in solving the EXPR and VA tasks. As a result, our method achieves the performance of 1.2542 on the validation set and 1.1640 on the test set, ranking 4th place in the MTL Challenge.