Learning Domain-specific Languages and Synthesizing Axiomatizations
31/10/2025, 11:00am
Speaker
Abstract
Symbolic languages and their abstractions are fundamental to many problems in computing, including programming, compilation, and code optimization, as well as program reasoning, testing, and verification. Learning symbolic concepts expressed in these languages thus has widespread application to automation, e.g., automated discovery of compiler optimizations, program invariants and specifications, and programs themselves, as well as exploration, e.g., automated theorem proving and conjecturing in math. Symbolic learning is especially powerful when the computations and concepts for the domain of interest have already been adequately represented in a domain-specific language (DSL) which constrains the hypothesis class to concepts relevant in that domain, thus permitting few-shot learning of its concepts. In this seminar, I will explain two branches of my work related to an emerging and underexplored problem in this space: the problem of automatically synthesizing DSLs. In the first half of the seminar, I will explain my work on learning DSLs in support of few-shot learning — with a detour to explain a technique for answering decidability questions in this setting — and, in the second half, I will explain my work on a problem related to learning symbolic languages in support of reasoning, namely, the problem of automatically synthesizing axiomatizations.
Bio
Paul is a postdoctoral fellow at Harvard and an incoming assistant professor in computer science at the University of Colorado-Boulder (Fall 2026). His PhD research developed formal connections between automata theory and program synthesis, formula learning, and learning domain-specific languages. His current interests include finding efficient algorithms for DSL learning as well as questions about how symbolic languages, and the abstractions they offer, might emerge through simple computational mechanisms that involve less engineering and design than is typical. Paul has a PhD in computer science from the University of Illinois.