ByteDance (字节跳动)
Sep 2026 — PresentInfrastructure - Chief Architect Office · Infra Agent RSI
Building an infrastructure data ontology and developing self-evolving decision-making agents.
I am a PhD student at the Institute of Automation, Chinese Academy of Sciences (CASIA) and the University of Chinese Academy of Sciences (UCAS). In 2024, I served as President of the Student Union of the School of Artificial Intelligence, UCAS. I am advised by Prof. Jian Cheng and Prof. Gang Li at CLab.
I work on efficient AI computing and hardware-software co-design: accelerating 3D Gaussian Splatting, optimizing LLM and VLA inference, and exploring AI infrastructure and agent systems.
I am currently an Infra Agent RSI in ByteDance’s Infrastructure - Chief Architect Office, working on infrastructure data ontology development and the self-evolution of decision-making agents. Previously, I worked on on-device VLA inference acceleration in Li Auto’s Foundation Models – VLA Model Architecture team.
Infrastructure - Chief Architect Office · Infra Agent RSI
Building an infrastructure data ontology and developing self-evolving decision-making agents.
Foundation Models · VLA Model Architecture · Research Intern
Worked on decoupled dual-system VLA inference and on-device acceleration for bandwidth-constrained chips.
Publications, preprints, and work under review. * denotes equal contribution.
MICRO, 2025.
GCC accelerates novel view synthesis from a reconstructed 3D scene by skipping unused Gaussians and reusing each Gaussian across image tiles.
ICLR, 2026.
arXiv, 2026.
From Splats to Silicon uses GPU profiling to trace when reducing work in 3D Gaussian Splatting actually saves memory traffic or shortens frame time.
Under Review, 2026.
GRADE schedules rendering work for static and dynamic Gaussian scenes. Its 4D implementation computes appearance only for candidates that survive filtering for the current camera.
MICRO, 2026.
DeGS speeds up high-resolution Gaussian rendering by grouping useful pixel work before blending, so more processing elements can run in parallel.
arXiv, 2026.
LATTICE schedules commands on an AI accelerator so that more data fits in on-chip memory, then refines the pipeline around the chosen memory layout.
MICRO, 2026.
arXiv, 2026.
DAC, 2026.
AAAI, 2026.
DATE, 2026.
DATE, 2026.
arXiv, 2022.
Institute of Automation, Chinese Academy of Sciences; University of Chinese Academy of Sciences.
Electronic Information Engineering.
Electrical Engineering and Computer Sciences (EECS).