Research

We study the foundations of learning and decision-making, connecting machine learning with optimization, game theory, and control.

AI evaluation and alignment

At Princeton AI Alignment and Safety (AIAS), we develop methods to evaluate capabilities beyond human expertise and align AI behavior through incentives and interaction.

• Measuring Intelligence Beyond Human Scale (SepaRank): model-generated challenges for relative evaluation without human-authored benchmarks or a semantic judge.
• AI Alignment via Incentives and Correction: a mechanism-design approach that adapts rewards for solver–auditor interactions using bandit feedback.

Online learning and optimization

Our work develops efficient algorithms for regret minimization, adaptive optimization, and decision-making under uncertainty.

• AdaGrad and adaptive regularization: adapting optimization updates to the geometry of past gradients.
• Online Newton Step: logarithmic regret for exp-concave losses through second-order online optimization.
• Sublinear-time optimization for linear classification and semidefinite programming.
• Projection-free optimization: online algorithms and linear convergence for smooth, strongly convex objectives over polyhedral sets.
• Matrix prediction and continuous multi-armed bandits.

For background, see Introduction to Online Convex Optimization and Optimization for Machine Learning.

Control and sequence modeling

We develop online nonstochastic control and spectral methods for learning dynamical systems and predicting long sequences. Our sequence-modeling work includes collaborations with Google DeepMind Princeton.

• Online nonstochastic control: learning control policies under adversarial disturbances and changing costs. The framework includes gradient perturbation controllers, logarithmic regret, unknown dynamics, and partial observation.
• Universal Sequence Preconditioning and its second-order extension: memory-efficient prediction using orthogonal polynomials.
• Universal Learning of Nonlinear Dynamics: provable prediction through observation-based spectral filtering, building on spectral filtering for linear dynamical systems and asymmetric systems.
• Spectral Transformers and FutureFill: efficient neural sequence architectures and fast generation for convolutional models.
• Learning and control: boosting for dynamical systems, maximum-entropy exploration, bandit feedback, black-box control, and adaptive control of time-varying dynamics.

Explore our featured projects, including SepaRank and recent sequence-modeling work. DeLuca and Venti are listed under earlier projects.

See Google Scholar for publications and our blog for research discussions.