Innovation in AI
What We Do
The Kempner Institute is dedicated to forging transformative AI capabilities that yield meaningful societal and technological impacts. We develop cutting-edge AI solutions, explore novel data sources, and delve into new application domains. Our recent work in this area includes advancing machine learning in resource-limited environments, particularly for on-device applications, navigating issues of copyright and watermarking in generative AI, and creating new opportunities for applications of ML architectures in bioengineering and drug design.
New Foundation Model Lets Robots Visualize How to Act
Kempner Institute Investigator Yilun Du and his team introduce Large Video Planner (LVP): a robot foundation model that uses video as the primary modality. It encodes physical dynamics and world semantics, providing robots with a rich prior for physical decision-making. LVP demonstrates that by training models to visualize how to act, researchers can leverage the vast corpus of human demonstration data. To support open and reproducible research in this direction, Du and his team have released the LVP model as well as the LVP-1M training dataset — a corpus of 1.4 million video clips — to the research community.
Research Projects
AI for Neurological Discovery
PROTON: A Relational Foundation Model for Neurological Discovery
Kempner researchers developed a relational foundation model for neurological discovery and evaluated it through discovery loops that connect AI predictions to experiments in Parkinson’s disease, bipolar disorder, and Alzheimer’s disease.
Efficient Optimization of Quantized Models
LOTION: Smoothing the Optimization Landscape for Quantized Training
A research team at the Kempner Institute developed LOTION, a principled framework for smoothing the discontinuous optimization landscape encountered in quantized neural network training by replacing the raw quantized loss with its expectation under randomized-rounding noise, enabling standard optimizers to converge reliably while preserving all global minima of the original problem.
Enhancing Algorithms for Reinforcement Learning
Accelerating RL for LLM Reasoning with Optimal Advantage Regression
A team led by Kempner Institute Investigator Kianté Brantley proposes a new RL algorithm that estimates the optimal value function offline from the reference policy and performs on-policy updates using only one generation per prompt.
