cv

S.M. (candidate, expected Dec 2027)

Harvard University

Computational Science and Engineering — GSAS / SEAS

• Quantitative Methods for Natural Language Processing
• Advanced Scientific Computing
• MLOps

B.S.

Peking University

Statistics — School of Mathematical Sciences

• Deep Learning and Reinforcement Learning
• Statistical Learning
• Stochastic Processes
• Mathematical Statistics

Research Intern, Qingyun Program

Tencent

Core contributor, UI-Mate technical report — the foundation GUI agent from Tencent's Hy Team.

• Owned task construction for OSWorkerBench, a benchmark of 100 long-horizon office workflows spanning 41 applications and 10 job families, released with UI-Mate.
• Designed the capability-grounded task generation pipeline and made every task automatically scoreable by weighted checkpoints on final application state, then refined them through iterative human-in-the-loop verification using pilot-agent rollouts and synthetic test cases.
• Recorded human demonstrations for hard tasks, making this the first computer-use benchmark to supply multimodal demonstrations as one-shot guidance and to test procedural transfer from related but non-identical source tasks.

Research Associate

Carnegie Mellon University, Machine Learning Department

Advised by Prof. Katerina Fragkiadaki.

• RobotArena ∞: unlimited robot benchmarking via real-to-sim translation (ICLR 2026). Built a real-to-sim pipeline converting teleoperation video into physics-consistent 3D environments via generative scene modeling and differentiable rendering.
• Designed automatic camera–robot calibration using 3D Gaussian Splatting, reaching sub-centimeter pose error with no manual supervision; applied it to BridgeV2, DROID and RH20T, obtaining more than 10k calibrated trajectories.
• ParticleDiffuser: generative 3D object-particle dynamics. Developed a diffusion model predicting 3D object-particle and gripper trajectories from point clouds, covering rigid and deformable objects.
• Built a goal-conditioned guided-diffusion controller with 3× lower MSE and 20× faster planning than an MPC baseline; extended guidance across prediction chunks for long-horizon planning, and validated sim-to-real transfer.

Research Assistant

Peking University, Wangxuan Institute of Computer Technology

Advised by Prof. Jiahuan Zhou, OV³ Lab.

• SCAP: supportive clique-based attribute prompting for test-time adaptation (CVPR 2025). Preserved vision-language model accuracy under deployment distribution shift with no labels and no retraining by adapting the model dynamically based on cliques of visually similar samples, achieving further improved performance on larger batches.
• D2ART: dual-modal prompting for vision–language models. Adapted CLIP at test time using entropy-guided prompt updates and EMA feature stabilization, improving top-1 accuracy by 2.3% on ImageNet-A over prior test-time adaptation methods.

Multi-Skill Dexterous Manipulation: RL to flow-matching

Advised by Dr. He Wang.

• Built a three-stage learning pipeline for dexterous grasping: a flow-matching pose proposer conditioned on object point clouds, an RL pick-up policy rewarded against proposed grasps to prevent mode collapse, and a visuomotor student distilled from RL demonstrations.
• Injected noise during demonstration collection to collect recovery behaviors from the teacher, addressing behavioral-cloning distribution shift; reached 85%+ success across Dexonomy with diverse grasp strategies.

Throughline: goal-tracking planner that turns notes and meetings into todos

Designed, built and deployed.

• Built and deployed a planning app tracking long- and short-term goals: a pre-filter narrows candidates before the model adjudicates each proposed item as new, duplicate, or an in-place update against the live task list.
• Designed the Profile that represents the user's background, goals and opinions, enabling the app to analyze notes and todos in the user's perspective to produce tailored advice. Implemented all app functions around the profile.

UI-Mate: Advancing Open-Weight Foundation GUI Agents with In-Context Demonstrations

Tencent Hy Team — technical report (arXiv:2608.15930)

Core contributor. Owned task construction for OSWorkerBench, the accompanying benchmark of 100 long-horizon office workflows.

RobotArena ∞: Scalable Robot Benchmarking via Real-to-Sim Translation

International Conference on Learning Representations (ICLR) 2026

Real-to-sim pipeline and automatic camera–robot calibration via 3D Gaussian Splatting, yielding 10k+ calibrated trajectories across BridgeV2, DROID and RH20T.

SCAP: Transductive Test-Time Adaptation via Supportive Clique-Based Attribute Prompting

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2025

First author. Label-free, retraining-free test-time adaptation of vision-language models using cliques of visually similar test samples.