Abstract: Bayesian optimization (BO) is a principled framework for efficiently optimizing expensive black-box functions. Its standard formulation typically assumes a fixed search space, an objective specified in advance, and observations of a predetermined form. In many real-world design problems, however, objectives are difficult to articulate, search spaces evolve as results arrive, and useful information often lies outside the optimization formalism. Large language model (LLM) agents can interpret this information, draw on domain knowledge, and use software tools. This talk explores how to combine that flexibility with probabilistic surrogate models and acquisition functions for uncertainty-aware, sample-efficient search.

I will present two recent projects. The first, LILO (Language-in-the-Loop Optimization), addresses optimization without an explicit objective. An LLM translates a decision maker’s free-form feedback into preference labels over observed outcomes, and a Gaussian process models the resulting latent utility to guide future evaluations. In our experiments, LILO outperforms preferential BO, LLM-only optimizers, and even standard BO with direct access to the true utility. A single natural-language message can also convey information comparable to more than a dozen pairwise comparisons.

The second project, agentic BO, places an LLM agent in charge of the optimization loop while a Bayesian backend maintains the surrogate model. Before each expensive evaluation, the agent can query the surrogate, accept or override its proposals, and modify the search region, acquisition function, objectives, or constraints. In our experiments, the system matches a well-tuned BO policy without prior domain knowledge and improves on it when given relevant natural-language knowledge. It can also reformulate a problem during an optimization campaign while retaining earlier observations.

Bio: Max Balandat is a Principal AI Engineer at Atomic Machines, where he leads research and development in AI-native design and optimization for micro manufacturing. His work combines generative design, agentic workflows, and optimization to advance the company’s “prompt-to-product” vision. Previously, he spent ten years at Meta, where he led the Modeling & Optimization group within Central Applied Science’s Adaptive Experimentation team. He created BoTorch, a PyTorch library for Bayesian optimization, and co-created Ax, a platform for adaptive experimentation and optimization. His research interests include probabilistic machine learning, uncertainty quantification, sample-efficient optimization, and their application to complex multiscale and multiphysics design problems. Max holds a PhD in Electrical Engineering and Computer Sciences and an MA in Mathematics from UC Berkeley.