Swarat Chaudhuri
University of Texas at Austin, TX·Computer Science
Neural networks are good at spotting patterns in messy data, while symbolic methods such as logic and formal programs are good at precise, checkable reasoning. Combining the two is called neurosymbolic AI.
- Fields
Computer Science
Mathematics & Statistics
- Topics
neurosymbolic programming
neurosymbolic AI
foundation models
- Grants
- 4 known grants, through Sep 2028
What they work on
- Develop foundation model programming, which writes scientific hypotheses as programs combining symbolic rules with neural modules and foundation models, to extract interpretable concepts from images and video.
- Create algorithms that jointly choose which interpretable concepts to extract from raw data and how to compose them into hypotheses, benchmarking them on tasks across several scientific domains.
- Train large language models on proof data and combine reinforcement learning with search to automate low-level steps in interactive theorem provers such as Coq and Lean.
- Discover new proof tactics, reusable proof steps, automatically through proof compression, treating formal theorem proving as a control problem.
- Build a neurosymbolic program-synthesis framework in which transformer code models learn explicit program semantics and guide specification-directed synthesizers to avoid errors like nonterminating loops.
- Develop a tunable, full-stack foundation for chain-forward declarative reasoning with probabilistic and higher-order features that runs at scale on supercomputers, GPUs and cloud clusters.
More about the research
Across four projects, this research uses that combination to make AI write correct code, prove mathematical theorems, generate understandable scientific hypotheses, and run logical reasoning efficiently on large computers.
Why it matters
More reliable AI-generated code could reduce hidden software vulnerabilities. Automating proofs could make formal verification practical for safer software, sturdier hardware and more rigorous mathematics. Interpretable hypothesis tools could help scientists in fields like neuroscience and ecology learn from images and video with less manual labeling. Scalable declarative tools could serve programmers in precision medicine, security and graph analytics.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Collaborative Research: III: Medium: Interpretable Scientific Discovery with Foundation Model Programs
NSF · Oct 2025 – Sep 2028
Collaborative Research: PPoSS: Large: A Full-stack Approach to Declarative Analytics at Scale
NSF · Aug 2023 – Jul 2028
SHF: Medium: Neurosymbolic Agents for Formal Theorem-Proving
NSF · Jun 2024 – May 2028
Closest labs
Binoy Ravindran
Virginia Polytechnic Institute and State University, VA·Computer Science
Develop methods that lift binary code into a provably over-approximate high-level model, one that captures both intended and unintended program behavior, for reasoning about exploitable flaws.
Xiaokang Qiu
Purdue University, IN·Computer Science·Early-career
Develop program synthesis algorithms that infer objective functions, library models and task decomposition themselves, so programmers need not hand-craft them, enabling scalable library-based synthesis.
Ravi Mangal
Colorado State University-Fort Collins, CO·Computer Science
Design quantitative metrics for evaluating code generated by large language models using symbolic execution, which explores program paths with symbolic inputs, and Bayesian program analyses that reason probabilistically about program behavior.
The school
University of Texas at Austin: 49 labs on PI Pages, $1.1B a year on research, 871 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.