MIT Class of 2028 - AI + Mathematics - Generative Models
I am an undergraduate at MIT, double majoring in Artificial Intelligence and Decision Making (Course 6-4) and Mathematics (Course 18). I work on generative modeling in He Vision Group, led by Kaiming He, as a UROP student, with a current focus on fast, simple, and principled image generation: diffusion and flow matching, normalizing flows, one-step generation, and multimodal understanding/generation.
One-step and low-NFE image generation in pixel space and latent-free settings.
Normalizing flows, diffusion, and flow matching with a focus on what structure is actually necessary.
Vision-language understanding and text-to-image generation under a shared modeling view.
ICML 2025 poster - first author
We revisit a common assumption in diffusion and flow-matching models: whether denoisers need explicit noise conditioning. Across eight reimplemented denoising generative models, uEDM shows that noise-unconditional diffusion can remain competitive, with theory matching the empirical behavior.
CVPR Spotlight - project lead - first author
BiFlow revisits normalizing flows with a learned reverse map guided by hidden-state alignment. It removes the need for explicit inverse-flow computation, avoids slow autoregressive inference, and enables single-evaluation NF-based generation with strong fidelity.
ICML accepted - first author - one-step pixel-space generation
pMF builds a strong baseline for one-step, latent-free generation by using MeanFlow with x-prediction directly in pixel space. The project reports 2.22 FID on ImageNet 256 and 2.48 FID on ImageNet 512.
Jun 20 - Aug 31, 2025 - quantitative research
Worked on quantitative strategy analysis, backtesting workflows, factor diagnostics, and reproducible Python research utilities for strategy evaluation.
MIT 6.4210 course project - robotics manipulation
Reimplemented a diffusion-based humanoid loco-manipulation pipeline inspired by BeyondMimic, then replaced DDPM sampling with flow matching. The project used simulation data, motion tracking, post-hoc control guidance, and lower-latency sampling with 5 FM steps.
Massachusetts Institute of Technology, undergraduate in AI and Mathematics, GPA 5.00/5.00.
Top 17 in the 2025 Putnam Mathematical Competition.
2nd place in the 2024 Putnam Mathematical Competition.
Gold Medal and 11th place at the International Mathematical Olympiad.
Pre-college student at Tsinghua University's Institute for Interdisciplinary Information Sciences, GPA 4.00/4.00.
Gold Medal and 1st place with perfect score in the Chinese Mathematical Olympiad.
Excellent Award in Alibaba Global Mathematics Competition, top 70 among 50,000+ participants.
