Research

My research develops reliable methods for causal inference and human-centered AI, with a focus on trustworthy automation for data science. It is organized around a single question:

How can we automate causal inference — from the statistical estimator to the AI agent — without giving up the mathematical guarantees, transparency, and domain sensitivity that scientific and policy applications require?

I pursue that question along two complementary lines. The theoretical line develops statistical and computational methods that automate parts of causal analysis which currently demand substantial expert judgment, while preserving rigorous guarantees. The empirical line studies how AI systems themselves reason about cause and effect, and builds tools to evaluate and formally verify that reasoning.

Partial identification: making uncertainty explicit

In many empirical settings, the data and assumptions are not strong enough to identify a single causal estimand, but they still imply informative bounds. I treat partial identification not as a last resort but as a principled way to report what the data can and cannot support: where a point estimate would over-promise precision, credible bounds tell decision-makers the truth.

Consistency of Neural Causal Partial Identification (NeurIPS 2024) gives a general computational approach to partial identification under structural causal constraints, proves consistency, and characterizes the approximation power of the model class.
Partial Identification of Policy-Relevant Treatment Effects with Instrumental Variables via Optimal Transport casts the problem as optimal transport and derives a tractable estimator with finite-sample guarantees, for settings where researchers need interpretable bounds under credible but incomplete assumptions.

Automating causal inference

Causal and econometric estimation is full of decisions that today rest on an expert’s intuition — how much to regularize, which estimator to trust, how to tune it. These choices can decide whether an estimator is stable or misleading, especially in ill-posed problems that recur across empirical economics, from demand estimation to program evaluation.

Adaptive Estimation and Inference in Conditional Moment Models via the Discrepancy Principle selects regularization parameters from the data in nonparametric instrumental-variable problems, with oracle inequalities and valid inference.
During a summer at LinkedIn, I prototyped a debiased machine learning framework for predicting long-term treatment effects when short-term A/B-test outcomes are noisy or delayed, and identified failure modes under confounding shifts.

Making these choices data-driven and theoretically controlled lowers the barrier between advanced causal methodology and reliable empirical practice.

How AI systems reason about cause and effect

The same question, turned on AI itself: is a model’s causal reasoning real, or only apparent? I study this at two levels.

At the level of behavior, CausalReasoningBenchmark builds a benchmark from published empirical studies that separates causal identification from causal estimation. The separation matters: a fluent causal explanation can still rest on the wrong identification conditions or target the wrong estimand.

At the level of internal mechanism, Bucketing the Good Apples studies causal abstraction — whether a model’s internal computation faithfully realizes a higher-level causal structure — and gives a method for diagnosing and repairing where that abstraction breaks down.

Ongoing: formally verified causal inference

My current work brings the two lines together. I am building a Lean-based library for causal inference, together with a research pipeline in which AI agents propose structured causal questions, develop candidate theorems, and submit them to formal verification. The rigor of the theory then becomes a guarantee the AI systems can be held to — machine-checkable, even as AI takes on more of the reasoning.

The aim is to let AI assist with literature organization, theorem discovery, and proof checking, while the human retains responsibility for judgment.

Other work

Some questions do not belong to any of the three threads above.

Treatment effects in the tails. Estimation of Treatment Effects in Extreme and Unobserved Data (NeurIPS 2025) asks what can be said about treatment effects in regions the data barely reach. Rather than bounding the effect, it uses extreme-value theory to extrapolate beyond the observed support, with consistency guarantees and validation on synthetic and real data.

Optimization and reinforcement learning. Before turning to causal inference, I worked on a homogenization approach for gradient-dominated stochastic optimization (UAI 2024), the derivative-free solver SOLNP+, and pessimistic minimax value iteration (ICML 2022), which characterizes the data coverage needed to learn Nash equilibria offline.