We strive for Efficient and Effective AI –
advancing the efficiency, reasoning, and creativity of AI models.
It aims to improve the efficiency of training (including fine-tuning) and inference for AI models like LLMs. Our work spans novel training algorithms/optimizers (e.g., Win, Adan), efficient inference techniques (e.g., GRIFFIN ), streamlined neural network architectures (e.g., PoolFormer), and advanced learning frameworks (e.g., FDM)—all designed to maximize speed and memory efficiency without compromising accuracy.
We focus on enhancing AI models' reasoning and collaboration capabilities through agentic reasoning and collaboration (e.g., CaPo and e.g., CoTS). Our approach empowers AI agents to plan dynamically, leverage world knowledge, reason across modalities using a lightweight global workspace, and coordinate with one another to integrate perception, memory, and planning for robust decision-making.
We develop AI models with controllable generation and editing capabilities to enhance their imagination and creativity. Our work, e.g., MDT, EditAnything, and Consistent3D, focuses on generating and editing high-quality image, 3D, and video content according to user’s commands like text, drag points and sketch. By enabling fine-grained control over the output, we empower AI to support tasks such as visual content creation, interactive design, and storytelling—bringing human-like creativity into generative models.
An ``aircraft'' AI model
empowered by Efficiency, Reasoning and Creativity