LLMs as counterfactual reasoners
Next-token training rewards plausible text, so a model can answer three equivalent counterfactual questions three different ways without ever paying a penalty.
Advanced Learning and Reasoning Lab
We develop learning frameworks for fundamental bottlenecks in modern AI, with particular emphasis on nonstandard targets in complex, unstructured, and often large-scale data. One such bottleneck is not data, compute, or model capacity, but the lack of a learning theory for questions about worlds that were never observed: what would happen under a different action, policy, or environment, and how reliably such alternatives can be learned from available evidence. These are the nonstandard targets we set out to learn — quantities that matter for a decision yet are never measured directly. Rather than treating familiar AI failures as isolated phenomena, we study them as one family of unresolved “what if?” questions and build methods that move beyond associational prediction while keeping statistical guarantees. This places our work at the intersection of AI transparency, safety, fairness, and robustness, aimed at reliable decisions under intervention, shift, and uncertainty.
Addressing these questions requires ideas that cut across fields such as statistics, mathematics, and computer science. Our projects range from foundational theory and methodology to applications in high-stakes domains such as healthcare, finance, marketing, public policy, and manufacturing. In our lab, each student is expected to take intellectual ownership of their work and lead the development of their own research papers.
Current Directions
Four representative projects currently underway in our lab, spanning counterfactual reasoning, evidence transport, safe decision-making, and scientific discovery
Clinical-trial extrapolation
A trial reports one average effect, but the same treatment can help one subgroup and harm another, so effects must be transported to the population a clinician actually treats.
Offline evaluation and safe deployment
Deciding whether to deploy a new policy requires estimating its value from data collected under a different one, exactly where the logged data is thinnest.
Scientific discovery as intervention
The space of possible perturbations is vast, only a small fraction can be observed, and experiments are often costly to run. Discovery therefore depends on predicting the outcomes of experiments never performed.
Student-Led Work
Topological causal effectsWe present a framework for capturing treatment-induced topological shifts in complex, non-Euclidean outcomes by introducing a novel topology-aware causal estimand, the topological average treatment effect (TATE), defined as the expected difference between the power-weighted silhouettes of potential outcomes. We develop an efficient, doubly robust estimator for this estimand in a fully nonparametric model, establish functional weak convergence, and construct a formal test of the null hypothesis of no topological effect. We empirically demonstrate the effectiveness of our estimator across diverse types of complex outcomes, including images, graphs, and point clouds.
International Conference on Learning Representations (ICLR), 2026
Provably correct k-means clustering of persistence diagramsWe consider surrogate k-means clustering of persistence diagrams via their Hilbert space summaries, most notably persistence landscapes. We provide geometric margin conditions under which the cluster assignments induced by the landscape codebook match those induced by the corresponding diagram codebook, and establish risk calibration results showing that the diagram-space excess risk can be controlled by the landscape-space excess risk while inheriting fast k-means rates from the Hilbert space. Empirical studies demonstrate that surrogate k-means clustering using persistence landscapes effectively recovers the true diagram codebook.
Conference on Uncertainty in Artificial Intelligence (UAI), 2026
Causal evaluation of early-stage fast-charging protocols under cell-to-cell variabilityA framework for fairly evaluating early-stage fast-charging protocols while accounting for cell-to-cell variability in lithium-ion batteries. We estimate which protocols cause the least battery degradation, and examine whether other protocols may be more effective for particular subgroups of cells given their initial conditions.
IEEE Transactions on Industrial Informatics, 2026
ECLayr: A Fast and Robust Topological Layer via Euler Characteristic CurvesWe propose ECLayr, an efficient topological layer for general deep learning models based on the Euler Characteristic Curve. We introduce a novel backpropagation strategy that improves computational efficiency and mitigates vanishing gradient issues of prior approaches, and establish explicit stability guarantees for the proposed layer in the presence of noise and outliers. We demonstrate the effectiveness of our approach through experiments on a variety of datasets, including high-dimensional settings where persistent homology-based layers become computationally prohibitive.
Under review
Doubly robust Q-learning for right-censored survival outcomes via AIPCW temporal-difference targetsWe develop a doubly robust Q-learning framework for estimating optimal dynamic treatment regimes with right-censored survival outcomes. The method constructs augmented inverse probability of censoring weighted temporal-difference targets that combine outcome regression and censoring models, and incorporates them into backward Q-learning while accounting for estimation error propagated across multiple treatment stages.
Under review
Profile-based kernel balancing for generalization and transportabilityWe develop a kernel balancing framework for generalization and transportability when individual-level data from the target population are unavailable and only aggregate summary statistics are observed. The proposed method constructs balancing weights using profile-based representations of the target distribution, and establishes conditions under which these summaries are sufficient for estimation.
Under review
Q-learning with Buckley–James imputation under right censoringWe study Q-learning for multistage treatment decisions when stage-specific survival outcomes and future-stage information may be censored. The proposed approach uses Buckley–James imputation for the current-stage reward and censoring-adjusted weighting for the future value to identify the Bellman target from observed data.
Under review
Learning under counterfactual shifts in image acquisitionThis work studies image classification for a counterfactual deployment setting in which the same population is assigned differently across image-acquisition environments, such as hospitals, devices, or imaging protocols. It considers how to learn an image-only classifier from partially labeled historical data without requiring data collected under the target deployment setting.
Under review
Will our treatment policy really work? Multiscale segmentation of treatment benefit regionsWe present a post-estimation diagnostic framework for assessing whether threshold policies based on estimated heterogeneous treatment effects are practically deployable. We characterize the structural simplicity, stability, and causal support of the resulting treatment-benefit region using persistent segmentation summaries and their overlap-trimmed analogues, together with the standard treatment coverage curve. We establish theoretical guarantees showing how persistent fragmentation constrains the complexity of policy classes, and prove that uniform accuracy in heterogeneous treatment effect estimation transfers to stable segmentation summaries.
Under review
Tracking Time-Resolved Phase Organization in Coupled Oscillator PopulationsThis project studies how phase organization in coupled oscillator populations evolves over time, moving beyond a single synchronization-onset summary. We use circular kernel density estimation to characterize multimodal phase structure and concentrated subpopulations at each time point, and track their continuation, split, merge, birth, and death using oscillator identity information. The framework is intended to provide an interpretable description of population reorganization in real oscillatory systems.
In preparation
Causal Clustering for Subgroup Discovery under Longitudinal Treatment RegimesA causal clustering methodology for identifying subgroups with heterogeneous responses to longitudinal treatment regimes. The difficulty is that an individual's response can depend on when treatment is assigned, so we cluster estimated counterfactual responses under different treatment timings to recover subgroups with genuinely distinct responses.
Under review
Sustained Engagement Strategies for Chronic Low Back Pain Using PROPEL Trial DataThis study estimates counterfactual mean outcomes for pain and disability under sustained engagement strategies involving yoga, nurse-delivered counseling, and mobile app use among adults with chronic low back pain. Using data from the PROPEL randomized trial, this secondary causal analysis applies g-computation, inverse probability weighting, and doubly robust estimation to address measured time-varying confounding.
Under review
Opportunities Open
2026.09
Undergraduate Research Intern
Recent Milestones
-
2026
Hajin Lee's paper on topological causal effects accepted at ICLR.
-
2026
Hajin Lee's persistence-diagram clustering paper accepted at UAI.
-
2026
JungHyun Kim's fast-charging paper accepted at IEEE Transactions on Industrial Informatics.
-
2026
Samsung Science and Technology Foundation grant awarded, funding the group through 2029.