CV

Education

Massachusetts Institute of Technology, Cambridge, MA
- B.S. in Physics / Artificial Intelligence and Decision Making; GPA: 5.00/5.00
- Aug. 2025 – Present
Tsinghua University, Institute for Interdisciplinary Information Sciences (IIIS), Beijing, China
- Preparatory Program and Freshman Year; GPA: 3.95/4.00
- Aug. 2024 – July 2025

54th International Physics Olympiad (IPhO): Gold Medal, 1st Place in Theoretical Round (July 2024)
9th Romanian Master of Physics (RMPh): Gold Medal, 3rd Place (March 2023)
HackMIT 2025: 2nd Place Modal Prize (Sept 2025)
MIT Informatics Tournament (MITIT) 2025 Winter Contest: 1st Place, Beginner’s Round (Dec 2025)
Tsinghua University: Xuetangban Scholarship & Freshman Scholarship (Dec 2024)

Undergraduate Researcher, Diffusion Models & Molecular Dynamics (Feb. 2026 – Present)
- MIT; Supervised by Prof. Tommi Jaakkola
- Researching the intersection of diffusion models and molecular dynamics.
Undergraduate Researcher, Reinforcement Learning & Embodied AI (Oct. 2025 – Present)
- FortyFive AI Lab, MIT; Supervised by Dr. Ge Yang
- Developing RL and imitation learning pipelines for humanoid whole-body control; investigating scaling laws for policy performance and generalization.
- Contributed to productionization of Vuer (3D visualization) and ML-Dash (experiment tracking & data storage).
Undergraduate Researcher, Learning-based Control (Mar. 2025 – Aug. 2025)
- Tsinghua University; Supervised by Prof. Huazhe Xu
- Investigated RIR (RL to Imitation to Real-world) framework for robot manipulation on Franka arm, leveraging PPO and DrQ-v2.
- Designed multi-stage approach for specialist training and multitask generalization via imitation learning and sim-to-real transfer.

Fast Humanoid Loco-Manipulation via Flow Matching (MIT 6.4210)

Compared DDPM and Flow Matching for humanoid control via trajectory synthesis; FM achieves 820 vs. 280 survival steps at 5-NFE.
Zero-shot loco-manipulation from walking-only data via test-time classifier guidance.
[Code] [Video]

RL vs. SFT for Mathematical Reasoning in LLMs (MIT 6.4610)

Compute-controlled comparison of PPO, GRPO, GMPO, RLOO against SFT on Qwen3-8B.
GMPO achieves 74.2% on GSM8K (vs. SFT 76.7%), demonstrating RL can match SFT without step-by-step supervision.
[Code]

Enhancing Diffusion Models with RL and Adversarial Rewards

Formulated reverse diffusion as MDP with adversarial discriminators, achieving 21.7% FID reduction vs. baseline. Plug-and-play for existing models.
[Code & Report]

PaperPlay: Hand-drawn Sketches to Playable Games (HackMIT 2025)

Built system turning sketches into games using OpenCV, physics engine, and real-time AI commentary.
[Demo] [Video]

Consistent Local Video Editing via Attention Manipulation (CLEVAM-DM)

Training-free framework for local video editing using BrushNet inpainting, DDIM inversion, and PerVFI for temporal coherence.
[Code & Report]

Daily Papers: Personalized ArXiv Research Digest

Agentic LLM pipeline for autonomous paper discovery: multi-step relevance filtering, ranking, and summarization over arXiv/HuggingFace feeds with automated email delivery.
[Code]

Algorithm Design for the Metric k-Center Problem

Authored survey and evaluation framework. Proposed algorithms achieving empirical approximation ratio 1.049 (vs. SCR 1.064).
[Code & Survey]

ML/AI: Symmetry & Applications to ML (6.7970/8.750), Advances in Computer Vision (6.8300), Robotic Manipulation (6.4210), Natural Language Processing (6.4610), Computer Vision (6.4300), Deep Learning (Tsinghua), Intro to Inference (6.3800), Intro to Machine Learning (6.3900)
Math/Theory: Quantum Computation (18.435), Quantum Physics II (8.05), Design & Analysis of Algorithms (6.1220), Mathematics for CS and AI (Tsinghua)