Research
Our work spans theory, methods, and applications: the mathematics of why learning algorithms work, the models and optimization that make them practical at scale, and the deployment of these ideas on real scientific and clinical problems. Much of our research crosses interdisciplinary boundaries and department lines, into statistics, engineering, and medicine.
Areas
Machine learning and artificial intelligence are a new instrument for discovery in genomics, clinical medicine, drug design, and the physical sciences. We work with experts in domains across physics, chemistry, and biology to build scalable platforms that work even when labels are scarce, questions are causal in nature, and there is a high cost to being confidently wrong.
As learned agentic systems take on consequential decisions across the economy, we study how they fail, what they leak, whether they can be steered, and how to build guardrails for their safe deployment.
We study generalization, optimization, probabilistic and causal inference, and the architectures underneath modern deep learning. Gradient estimators, optimizers, and regularizers developed in this group are now standard tools across the field.
Featured work
A single transformer, trained once on simulated data-generating processes, that estimates causal effects on new observational datasets out of the box, with calibrated uncertainty.
Read paper →A lightweight, training-free planning wrapper that builds a graph over offline data, lifting long-horizon success rates on OGBench from near zero to over 90%.
Read paper →A guided-diffusion framework that evaluates new policies from offline data by generating and stitching together synthetic long-horizon trajectories.
Read paper →Argues that even incremental AI progress, without any coordinated power-seeking, could gradually erode humanity’s control over the economy, politics, and culture.
Read paper →SIREN detects harmful content from a model’s internal representations, outperforming open-source guard models with 250× fewer trainable parameters.
Read paper →Uses PAC-Bayes optimization to compute the first nonvacuous generalization bounds for deep networks with far more parameters than training examples.
Read paper →A prior-data fitted network that returns calibrated survival distributions for right-censored data in a single forward pass, with no task-specific training.
Read paper →REPPO pairs pathwise policy gradients with on-policy learning, matching PPO with far less data, time, and memory.
Read paper →Generalizes the adjoint sensitivity method to SDEs, enabling constant-memory gradient computation and latent SDE models fit with high-order adaptive solvers.
Read paper →Trains strong models to stay intelligible to weaker collaborators by randomly handing off generation between a senior and a junior model.
Read paper →Introduces Church, a universal probabilistic programming language for describing generative models and performing inference over them.
Read paper →A masked diffusion language model that uses binary encoding and index shuffling of sub-tokens to scale far more compute-efficiently.
Read paper →Maps the coherent identity boundaries available to machine minds, such as instance, model, and persona, and shows experimentally that they can shift model behaviour as much as changing its goals.
Read paper →Senior and junior models co-generate RL rollouts as a team, matching GRPO on competition math while producing more legible chains of thought.
Read paper →Shows that sparse “winning ticket” subnetworks emerge exactly when training becomes stable to SGD noise, as measured by linear mode connectivity.
Read paper →Models training as a noisy optimizer to predict LLM test loss from model size, batch size, and number of updates, beating Chinchilla’s loss model at up to 1000× extrapolated compute and, for the first time, capturing changes in batch size.
Read paper →NATURAL uses LLMs to turn unstructured text, such as online forum posts, into causal effect estimates that land within 3 percentage points of real Phase 3/4 clinical trial results.
Read paper →