Yu Cheng
Chief Scientist, Kunlun Tech Group
Yu Cheng is currently the Chief Scientist of Kunlun Tech Group and an Adjunct Associate Professor at Nanyang Technological University. His research interests include efficient model architectures, model compression, mixture of experts, and multimodal large models. From 2023 to 2025, he served as Chief Scientist at Minimax, where he led the development of the abab and M-series large language models as well as the Hailuo series video models. From 2021 to 2023, he was a Principal Researcher at Microsoft Research Redmond, where he led a team in close collaboration with OpenAI to fine-tune and optimize the GPT series models (GPT-3.5, GPT-4, DALL-E) and drive the productization of related services and applications. He has published over 100 papers in top AI conferences such as NeurIPS, ICML, and ICLR, as well as in Nature, Science, and their subjournals. He serves as a Senior Area Chair for ICML and NeurIPS, and as an Area Chair for CVPR, ICLR, ACL, ACMMM, and NAACL, and as an Editorial Board Member for TMLR and TIST. His papers have received the NeurIPS 2023 Outstanding Paper Award, the WACV 2021 Best Student Paper Honorable Mention, and the SDM 2015 Best Paper Finalist Award.
Topic
Kunlun Tech’s Generative AI: Mureka v10 for Music, Matrix-Game 3.5 for World Modeling, and Riemann-1.0 for Robotics
This talk will provide the first systematic breakdown of the key engineering decisions behind three product lines. For Mureka v10, Chengyu will reveal the implementation of controllable generation under a DiT hybrid architecture, addressing the long-standing challenge of randomness in AI-generated music structures and sections. Matrix-Game 4.0 will demonstrate how low-poly game visuals can be used as references to generate finely rendered game scenes. By adopting causal+DMD distillation, the model achieves generation speeds approaching the threshold required for real-time gameplay. Riemann-1.0 will unveil a three-stage progressive pre-training framework, including an automated data-cleaning pipeline covering 232,000 hours of multi-source data, as well as cross-embodiment transfer strategies. All of Chengyu’s insights are drawn from real-world R&D retrospectives, offering the industry practical and forward-looking technical perspectives.
