Agentic intelligence · Generative systems
I am a Ph.D. researcher at HKUST(GZ), advised by Prof. Lei Zhu. My research connects agentic post-training, recursive self-improvement, and long-horizon planning and harness design. Before that, I received my M.S. in Artificial Intelligence from Tianjin University.
My research asks how agents can sustain open-ended work, learn from execution, and improve the systems around them:
Multi-agent coordination is the key to reliable long-horizon execution: specialist agents decompose and coordinate evolving workflows, while shared context and harnesses preserve continuity across extended tasks in JarvisHub and Seedance Studio.
Auto research and recursive self-improvement (RSI) through the joint evolution of skills, trajectory data, evaluators, and harnesses. I apply this framework in SeedanceGame to iteratively improve its agent system.
Building multi-agent systems and harnesses for long-horizon workflows, with persistent project context, traceable execution, and coordinated branching. Exploring recursive self-improvement (RSI) through trajectory data, evaluators, and iterative harness evolution.
Built video–action data pipelines and action-conditioned world models, including progressive denoising and distillation for minute-scale autoregressive generation.
Developed AgentShot and MSBench, large-scale V2V lip-sync generation, and RL-based human-centric video captioning with FactorizedGRPO.
Models multi-shot creation as a sequence of agent decisions across plot, storyboard, visual anchors, and motion. Grid rewards, CoT-SFT, online planning experience distillation, and DPO improve policy quality and cross-stage consistency.
Uses a visual canvas as the user workspace, shared project state, and external memory. Typed nodes, traceable actions, and a protocol bridge support inspectable and recoverable long-running work.
Introduces MSBench and a five-stage, nine-agent framework for prompt understanding, storyboard planning, and consistent multi-shot generation using open-source video models.