I am an undergraduate at MIT, studying AI and mathematics. I am currently a Student Researcher on the Foundational AI Research team at Google DeepMind in Cambridge, MA. I work with my host, Deqing Sun, on optimization for recurrent methods and looped models.
Since 2024, I have also been an undergraduate researcher in the He Vision Group at MIT, advised by Kaiming He. There, I work on generative modeling and fast image generation.
I seek to understand how learning and computation give rise to intelligence, and what the next level of intelligence could be. My methods and topics have changed before and will change again, as they should. I hold them loosely, and hold firmly to the curiosity I started with.
Ongoing research on optimization for recurrent methods and looped models.
ICML 2025
Revisits whether diffusion and flow-matching models need noise conditioning, reimplementing eight denoising generative models across several datasets. Proposes uEDM, a noise-unconditional diffusion model with competitive performance, with theory that matches the observed behavior.
CVPR 2026Highlight
Revisits normalizing flows with a learned reverse map, removing explicit inverse-flow computation and slow autoregressive inference. Hidden-state alignment guides the reverse model, enabling single-evaluation generation with state-of-the-art results among normalizing flows.
ICML 2026
A strong baseline for one-step, latent-free image generation in pixel space, using MeanFlow with x-prediction. Reaches 2.22 FID on ImageNet 256×256 and 2.48 FID on ImageNet 512×512.
* Equal contribution · † Project lead
Reimplemented a BeyondMimic-style humanoid loco-manipulation pipeline with simulated data preparation, diffusion training, and post-hoc control guidance. Replacing DDPM sampling with flow matching gave lower-latency control with 5 sampling steps.
Top17
Putnam Mathematical Competition
2nd
Putnam Mathematical Competition
11th
International Mathematical Olympiad
1st
Chinese Mathematical Olympiad
Top70
Alibaba Global Mathematics Competition
