Surbhi Goel
University of Pennsylvania, PA·Computer Science·Early-career
Artificial intelligence systems increasingly help with decisions in medicine, science and education. Yet they still make simple reasoning mistakes.
- Fields
Computer Science
Mathematics & Statistics
- Topics
machine learning theory
AI reliability
AI safety
- Grants
- 2 known grants, through May 2031
What they work on
- Characterize why transformers learn brittle shortcuts rather than robust algorithms, using controlled mathematical environments that isolate real-world reasoning failures and remain tractable for rigorous analysis.
- Develop principled interventions in training data, training algorithms and inference, such as representative training sets and sample-efficient multi-step reasoning methods, to make transformer reasoning reliable and resistant to adversarial manipulation.
- Build a theory of human-AI collaboration when each side has different information and goals are imperfectly aligned, producing auditable protocols grounded in calibration (confidence that matches accuracy).
- Design efficient algorithms for Testable Learning with Distribution Shift (TDS learning) that output a classifier plus a proof that it is either optimal on the test set or facing significant distribution shift.
- Extend PQ learning, which lets a model abstain on individual shifted examples, and study sequential prediction on possibly adversarial data streams, using tools from pseudorandomness, sum-of-squares proofs and robust statistics.
More about the research
They can also fail badly when the data they meet in use differs from the data they were trained on. The research asks why AI models fail and how to design methods, backed by mathematical proof, that make them dependable. That includes models working on their own, models working alongside people, and models facing unfamiliar data. Both projects combine rigorous theory with targeted experiments, and both check predictions against real benchmarks.
Why it matters
AI tools are moving into high-stakes settings such as medical diagnosis, healthcare and autonomous systems, where wrong predictions can cause serious harm. This work aims to provide the scientific foundations for deploying AI more safely in these settings. It targets models that reason reliably, flag unfamiliar data instead of guessing, and collaborate with humans in auditable ways. The project will also release open-source evaluation tools for other researchers.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Cong Ma
University of Chicago, IL·Mathematics & Statistics·Early-career
Develop robust ranking algorithms for pairwise comparison data that tolerate ill-conditioned sampling of which pairs are compared and relax parametric modeling assumptions.
Paul Valiant
Purdue University, IN·Computer Science
Design error-correcting codes that withstand random or adversarial insertion and deletion errors while supporting local decoding, which recovers part of a message by reading only a small piece of the stored data.
Yogatheesan Varatharajah
University of Minnesota, MN·Computer Science·Early-career
Develop an attention-based EEG foundation model, shaped by knowledge of brain signals, that captures how activity changes across recording channels and over time.
The school
University of Pennsylvania: 90 labs on PI Pages, $2.2B a year on research, 750 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.