Research Thrust 1: Foundations
This thrust develops the core principles for optimizing AI models to learn, reason, and generalize more effectively and efficiently. It spans adaptive post-training and test-time optimization, efficient representation learning and model compression, out-of-distribution generalization in structured domains, and trustworthy intelligence through methods such as activation steering and ecosystem-wide watermarking.
TILOS Publications on Google Scholar (opens in a new tab)
Theme F1Adaptive Intelligence
Mathematical foundations and principled methods for post-training and test-time optimization, so that AI systems keep learning reliably beyond their initial training.
Research directions
- Reinforcement learning for post-training and test-time optimization
- Robustness, reasoning, and the prevention of catastrophic forgetting
- Guaranteed reliability via everywhere learning
Faculty
Alexander Rakhlin, Hamed Hassani, Tara Javidi, Alejandro Ribeiro, Yu-Xiang Wang
Project highlights for this theme are coming soon.
Theme F2Efficient Intelligence
Reducing the computational, memory, and energy costs of AI across the entire pipeline through more efficient learning algorithms and compression of data representations and models.
Research directions
- Foundations of LLM quantization and low-precision training
- Long-context learning for efficient intelligence
- Backpropagation-free learning for efficient AI
Faculty
Shirin Saeedi Bidokhti, Mikhail Belkin, Tara Javidi, Arya Mazumdar, Yury Polyanskiy
Project highlights for this theme are coming soon.
Theme F3Generalizable Intelligence
Understanding and improving how trained models perform on unseen data, including out-of-distribution settings and structured data such as sets and graphs.
Research directions
- Optimization-induced structure, stability, and generalization
- Out-of-distribution generalization in structured domains
Faculty
Yusu Wang, Hamed Hassani, Arya Mazumdar, Alejandro Ribeiro, Yu-Xiang Wang
Project highlights for this theme are coming soon.
Theme F4Trustworthy Intelligence
Rigorous, theoretically grounded methods for safe and verifiable AI as foundation models move into high-stakes, real-world deployments.
Research directions
- Precise control via activation steering
- Human–AI interaction and trustworthy deployment
- Watermarking for ecosystem-wide trust
Faculty
Hamed Hassani, Mikhail Belkin, Tara Javidi, Farinaz Koushanfar, Nisheeth Vishnoi, Yu-Xiang Wang
Project highlights for this theme are coming soon.
Theme F5Foundations of Agentic Intelligence
Collaborative learning and optimization among multiple agents working toward shared goals under communication and compute constraints.
Research directions
- Foundations of multi-agent learning and optimization
- Implicit bias of distributed learning algorithms
Faculty
Arya Mazumdar, Tara Javidi, Yian Ma, Shirin Saeedi Bidokhti
Project highlights for this theme are coming soon.
Theme F6Generative & Learning-based Optimization
Using generative models as scalable optimization primitives through a generate–steer–refine framework, with applications across reasoning, robotics, chip design, and systems.
Research directions
- Efficient, steerable, and reasoning-capable diffusion models
- Constraint-aware generative optimization
- Diffusion priors for multi-objective optimization
Faculty
Yian Ma, Sicun Gao, Melvin Leok, Alexander Rakhlin, Alejandro Ribeiro, Nisheeth Vishnoi, Yusu Wang
Project highlights for this theme are coming soon.