Skip to main content
Research topic

Neurosymbolic AI Research and Program Synthesis

Dr. Costilla Reyes’s work connecting machine learning, symbolic reasoning, program synthesis, and interpretable scientific models.

Combining learned representations with symbolic structure

Dr. Costilla Reyes’s neurosymbolic work studies how statistical learning can work alongside programs, reusable concepts, equations, and domain knowledge. Neural networks learn patterns from data but are hard to read from the inside; symbolic programs state their steps explicitly. Across these projects, the goal is to expose useful structure in a form researchers can inspect and revise, rather than treating every learned result as an opaque prediction.

Neurosymbolic Programming for Science describes combining learned representations with prior scientific knowledge, experimental context, high-level concepts, and known constraints. It uses behavior analysis to discuss scientific workflows and the challenge of keeping their outputs interpretable. The same research community organized the Neurosymbolic Programming Summer School at Caltech in July 2022 and the AI for Code and Science Workshop at MIT in February 2023, both of which Dr. Costilla Reyes helped organize.

From reusable abstractions to equation discovery

LEMMA turns previously discovered mathematical solutions into higher-level actions that can be reused on later problems. In experiments on equation solving and fraction simplification, it solved more problems and generalized more effectively to examples harder than those seen during training. MathDSL explores a related idea through a domain-specific language, a small purpose-built set of operations, for linear-equation solving; its evaluation considers solution conciseness as well as accuracy.

InceptionSR applies reuse to symbolic regression, the search for mathematical equations that fit data: useful partial equations become features in later search rounds. A learned concept library explores reusable abstractions for equation discovery, while other work compares evolutionary methods for producing interpretable models. These methods differ, but each makes the structure of a candidate solution part of the research output.

Interpretable models and expert review

A clinician-evaluated preprint on body-fat estimation paired symbolic-regression formulas with language-model explanations and rankings for review by a panel of clinicians, who rated the comparative rankings more favorably than the explanations of single equations. The study shows how comparative rankings can support expert review of candidate physiological explanations.

Together, this work advances reusable mathematical abstractions, scientific program synthesis, and equation discovery across a range of research tasks.

Topics