Co-Op, Bayesian Methods for Complex Decision Making

About the role#

Lila Sciences builds AI systems to accelerate discovery in the physical and life sciences. The Physical Sciences AI team creates autonomous discovery agents that determine which experiments our AI Science Factories should run. Scientific campaigns often involve complex constraints, mixing discrete choices like components with continuous settings like temperature. Because every experiment consumes time and materials, we need methods that go beyond off-the-shelf solutions. As a Co-Op, you will work with research scientists and engineers to improve how AI systems make experimental decisions. You might build agents that propose the next experiment, test the reliability of in-context learners, or develop Bayesian strategies for limited budgets. Your work will be tested against live campaigns.

What you'll do#

• Design and benchmark experiment-selection strategies that balance exploration and exploitation under realistic budgets.
• Build autonomous discovery agents that propose the next experiment from surrogate models.
• Develop in-context learning approaches inspired by Bayesian inference while evaluating calibration and robustness.
• Run retrospective and live evaluations on active campaign data using clear baselines and metrics.
• Present findings to ML and experimental teams and contribute to a publication.

What you'll need#

• You are pursuing a Master's or PhD in statistics, ML, CS, operations research, applied math, or a quantitative science.
• You have a solid grounding in Bayesian inference and decision making, including priors and posterior predictive distributions.
• You have research experience in Gaussian processes, Bayesian optimization, bandits, reinforcement learning, or probabilistic modeling.
• You possess strong Python skills with PyTorch and probabilistic libraries such as Pyro, PyMC, or BoTorch.
• You are comfortable working with small, noisy, and heterogeneous datasets.
• You can communicate clearly with both ML and experimental collaborators.
• Experience applying these methods to real problems in materials science, chemistry, or another physical science is a plus.
• Familiarity with in-context learning as implicit Bayesian inference, meta-learning, or prior-data fitted networks is helpful.
• We value experience in optimization over categorical or combinatorial spaces, reinforcement learning for sequential decision making, or training models on multimodal scientific data.

Location & details#

• This is a full-time, paid position.
• The role is on-site in Cambridge, Massachusetts or San Francisco, California.
• The term is rolling.