CVPR 2026 模型适应性研究盘点：从保留旧知识，到适应真实世界

Notable works include Quantum-Gated Task-interaction Knowledge Distillation for class-incremental learning, achieving competitive accuracy on benchmarks like CIFAR-100, and the Large-Scale Codec Avatars framework, enhancing 3D digital human modeling through extensive pre-training. These advancements aim to ensure AI models retain old knowledge while effectively adapting to new tasks and diverse data environments.

Key Points

• Quantum-Gated Task-interaction Knowledge Distillation improves class-incremental learning, reducing catastrophic forgetting.
• Large-Scale Codec Avatars framework enhances 3D digital human modeling with high fidelity and generalization.
• FEAT method stabilizes federated learning by aligning geometric structures across clients.
• PolyV framework enables cross-vision synergy, improving performance on image, video, and 3D tasks.
• Research reflects a transition from capability expansion to capability management in large model development.