Hi! I am a PhD candidate in the CILVR lab at NYU Courant advised by Mengye Ren and supported by the NDSEG fellowship. I am also a visiting researcher at Meta FAIR.
My goal is to build AI systems that can solve complex tasks in the real world to amplify human productivity and well-being. I have worked on multimodal LLMs [1], visual pretraining [2, 3, 4], world models [5, 6], and reinforcement learning [7]. I am currently interested in exploration and long-horizon training.
I interned at Meta FAIR working on multimodal pretraining and computer use agents. Before starting my PhD, I worked on the systematic equities research team at The Voleon Group. I completed my joint BS/MS in computer science at the University of Michigan. There, I was fortunate to work with Honglak Lee on reinforcement learning and Michael P. Wellman on multi-agent systems.
|
|
|
November 2026 |
Unbundling Multimodal Vision Encoding is released. |
|
September 2026 |
I joined Meta's AI Mentorship Program as a visiting researcher. |
|
April 2026 |
I started my internship at Meta FAIR working on multimodal and unified pretraining mentored by Jakob Verbeek and Hu Xu. |
|
January 2026 |
Midway Network is accepted to ICLR 2026. |
|
May 2025 |
I started my internship at Meta FAIR working on computer use agents mentored by Joseph Tighe and Pierluca D'Oro. |
|
February 2025 |
PooDLe is accepted to ICLR 2025. |
|
|
|
We disentangle the three design choices that vision encoders bundle in multimodal LLMs and find that what matters is dedicated vision capacity, which can be trained jointly from scratch and placed inside the early layers of the backbone. |
|
|
Sliced variance regularization learns smooth latent maps for planning, and slow channels with a gated residual predictor keep the task-relevant content. |
|
![]() |
ICLR 2026
Midway Network is a new self-supervised learning architecture that is the first to learn visual representations for both object recognition and motion prediction solely from natural videos by leveraging latent dynamics. |
|
CVPR 2026 ActiVis Workshop
Our method Canonical pretrains visual representations by modeling camera viewpoint change as an SO(3) rotation in latent space, allowing visuomotor policies to generalize to unseen camera angles without calibration. |
![]() |
ICLR 2026 2nd Workshop on World Models
LaMo is a structured latent dynamics model that enables long-horizon prediction by autoregressively predicting latent motion tokens used to advance a visual scene's latent state over time. |
![]() |
ICML Tokenization Workshop 2025
We introduce Discrete-JEPA, which learns discrete semantic tokens for improved symbolic reasoning and long-horizon planning. |
|
ICLR 2025
We propose a self-supervised learning framework that combines pooled and dense objectives to learn representations with spatial and semantic understanding from naturalistic videos. |
|
NeurIPS 2021 · NeurIPS Workshop on Deep Reinforcement Learning 2020
We leverage successor features to formulate a graph-based planning framework and goal-conditioned policy, enabling long-horizon goal-reaching in visual environments. |
![]() |
Games 2021
We use an agent-based model and empirical game-theoretic analysis to study price manipulation in financial markets and propose mitigation mechanisms. |
![]() |
ICAIF 2020 · ICML Workshop on AI in Finance 2019
We design learning-based trading strategies that have improved robustness to market manipulation and evaluate them with agent-based simulation. |
