Semantic Tube Prediction: Beating LLM Data Efficiency with JEPA
A JEPA-style regularizer that improves signal-to-noise ratio and preserves diversity during LLM fine-tuning.
Research
Selected papers and technical projects on Physical AI, interaction understanding, representation learning, and foundation models.
LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures
A JEPA based solution for LLMs that outperforms the standard LLM training objectives and is robust to overfitting.
Adversarial Score identity Distillation: Rapidly Surpassing the Teacher in One Step
A one-step adversarial distillation method for diffusion models that improves both sample quality and distillation efficiency.
Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation
A data-free method that distills the generative capabilities of pretrained diffusion models into a single-step generator.