Research

My research grows out of building AI systems that must work in practice. Across language models and Physical AI, I study how learning objectives and architectural constraints can help models form representations at the right level of abstraction—beyond individual tokens or objects, but more structured than entire scenes. The goal is to learn meaningful state that supports reliable prediction and action.

Predictive Representation Learning at the Right Granularity

From JEPA-style regularization to variable-granularity concepts in language models

#LLM #JEPA #representation learning

My initial work showed that JEPA-style prediction can be used as a regularizer alongside a primary learning objective, removing the need for a moving-average target network. This led to a broader question: rather than predicting between fixed views or individual tokens, can models learn the appropriate granularity of abstraction? Subsequent work explores randomized fractional views, boundary bottlenecks, and continuous concepts as mechanisms for concentrating information into representations of meaningful structure.

Semantic Tube Prediction: Beating LLM Data Efficiency with JEPA

ICML 2026 · Paper

First author · Co-authored with Yann LeCun

LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures

ICLR 2026 · Paper

First author · Co-authored with Yann LeCun

Current Direction: Developing a unified account of perception and generation as prediction at different levels of abstraction.

Read more
Interaction Understanding for Physical AI

Learning representations of meaningful relationships and state changes in the physical world

#Physical AI #interaction tracking

Mainstream perception systems identify objects or summarize entire scenes, but many real-world decisions depend on the relationships and state changes between them. This program studies interaction as the intermediate abstraction: grounded enough to reflect what is physically happening, yet structured enough to support prediction and action.

Current Direction: Learning interaction representations from largely unlabeled video and translating them into reliable, real-time perception models for new environments.

Read more
Earlier Research: Efficient Generative Modeling

One-step diffusion generation through score and distribution distillation

#diffusion models #model distillation #one-step generation

My earlier research focused on efficient generative modeling, developing score-based distillation methods for training one-step diffusion generators, including SiD and SiDA.

Adversarial Score identity Distillation: Rapidly Surpassing the Teacher in One Step

ICLR 2025 · Paper

Senior author

Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation

ICML 2024 · Paper

Senior author

Read more
Research Leadership and Trustworthy AI

I also lead collaborative research projects beyond my primary research programs, helping shape the research question, technical direction, evaluation, and presentation.

LoRA Users Beware: A Few Spurious Tokens Can Manipulate Your Finetuned Model

Spotlight, ICLR 2026 Workshop On Principled Design for Trustworthy AI · Paper

Senior author

Read more