Assistant Professor @ UPenn Wharton

Contact
yjinstat at wharton.upenn.edu
Room 441, Academic Research Building
265 South 37th Street, Philadelphia, PA

Research areas

This page organizes my main research areas. See the list of papers by year here.
(* equal contribution or alphabetical order)

1. Trustworthy AI & uncertainty quantification

When AI models are deployed in high-stakes scenarios, the challenges are not just about accuracy. When can their outputs be trusted? What guarantees can we provide for cases most relevant to the problem at hand? How could uncertainty quantification guide decisions and resource allocation? My research develops statistical methods with novel guarantees to answer these questions.

Selective conformal inference: identify instances worth acting on (e.g., active drugs, reliable LLM outputs) based on imperfect predictions, or report uncertainty quantification after data-driven decisions.
Conformal prediction in complex settings such as graphs, multiple environments and unmeasured confounding.
Uncertainty-guided decisions: connect UQ to decisions, operationalizing the guarantees into trust/defer policies and real-world resource allocation pipelines with foundation AI models.

Representative papers:

Optimized Conformal Selection: Powerful selective inference after conformity score optimization
Tian Bai and Ying Jin, 2024
Conformal alignment: Knowing when to trust foundation models with guarantees
Yu Gui*, Ying Jin* and Zhimei Ren*. Conference on Neural Information Processing Systems (NeurIPS), 2024
Confidence on the focal: Conformal prediction with selection-conditional coverage
Ying Jin* and Zhimei Ren*. Journal of the Royal Statistical Society: Series B (JRSS-B), 2025
Selection by prediction with conformal p-values
Ying Jin and Emmanuel Candès. Journal of Machine Learning Research, 2023

2. Inference in AI for science

Generative AI introduces a new paradigm for scientific discovery, but it also raises a central challenge: how do we ensure generated outputs are relevant, testable, and worth pursuing? I develop statistical frameworks and agentic workflows that automate parts of the scientific loop, from hypothesis generation to validation.

Representative papers:

ConfHit: Conformal generative design via nested testing
Siddhartha Laghuvarapu, Ying Jin and Jimeng Sun. International Conference on Learning Representations (ICLR), 2026
Automated hypothesis validation with agentic sequential falsifications
Kexin Huang*, Ying Jin*, Ryan Li*, Michael Li, Emmanuel Candès, and Jure Leskovec. ICML, 2025

3. Robust causal inference

Causal insights are most valuable when they transfer to new populations: policymakers care about whether the effect in an earlier experiment generalizes to new places, and whether new decision rules learned from historical data deliver desired outcomes in new populations. My research in this topic focuses on the key issue of distribution shift in these problems:

Generalizability, replicability: empirical distribution shifts in treatment effect generalization and methods to address them.
Robust and efficient causal inference via covariate adjustments and sensitivity analysis.
Policy learning with limited overlap: pessimism principles for efficient learning of data-informed personalized decisions.

Representative papers:

Beyond reweighting: On the predictive role of covariate shift in effect generalization
Ying Jin, Naoki Egami, and Dominik Rothenhäusler. Proceedings of the National Academy of Sciences (PNAS), 2025
Cross-balancing for data-informed design and efficient analysis of observational studies
Ying Jin and José R. Zubizarreta, 2025
Policy learning “without” overlap: pessimism and generalized empirical Bernstein’s inequality
Ying Jin*, Zhimei Ren*, Zhuoran Yang, and Zhaoran Wang, 2022. Annals of Statistics, 2025
Is pessimism provably efficient for offline RL?
Ying Jin, Zhuoran Yang, and Zhaoran Wang. Mathematics of Operations Research, 2024+. Short version in ICML 2021