The paper introduces the Dexterous Grasp Transformer (DGTR), a novel discriminative framework for generating diverse and high-quality dexterous grasp poses from object point clouds. DGTR formulates the grasp generation task as a set prediction problem and uses a transformer-based model to predict multiple grasp poses in a single forward pass. To address the optimization challenges in dexterous grasping, DGTR proposes two progressive strategies: Dynamic-Static Matching Training (DSMT) for stable training and Adversarial-Balanced Test-Time Adaptation (AB-TTA) for refining grasp poses during testing. DSMT enhances optimization stability by guiding the model to learn appropriate targets through dynamic matching and optimizing object penetration through static matching. AB-TTA improves grasp quality by minimizing a pair of adversarial losses, one repelling the hand from the object and the other attracting it towards the object's surface. Experimental results on the DexGraspNet dataset demonstrate that DGTR outperforms previous methods in both grasp quality and diversity, achieving high-quality grasps with minimal object penetration and a wide range of grasping directions. The code and project page are available online.The paper introduces the Dexterous Grasp Transformer (DGTR), a novel discriminative framework for generating diverse and high-quality dexterous grasp poses from object point clouds. DGTR formulates the grasp generation task as a set prediction problem and uses a transformer-based model to predict multiple grasp poses in a single forward pass. To address the optimization challenges in dexterous grasping, DGTR proposes two progressive strategies: Dynamic-Static Matching Training (DSMT) for stable training and Adversarial-Balanced Test-Time Adaptation (AB-TTA) for refining grasp poses during testing. DSMT enhances optimization stability by guiding the model to learn appropriate targets through dynamic matching and optimizing object penetration through static matching. AB-TTA improves grasp quality by minimizing a pair of adversarial losses, one repelling the hand from the object and the other attracting it towards the object's surface. Experimental results on the DexGraspNet dataset demonstrate that DGTR outperforms previous methods in both grasp quality and diversity, achieving high-quality grasps with minimal object penetration and a wide range of grasping directions. The code and project page are available online.