Symbolic Regression with Self-Supervised Heuristic Beam Search

ICLR 2026 Conference Submission19041 Authors

Keywords: symbolic regression, self-supervised, heuristic search, neural network, deep learning

TL;DR: A new method for Symbolic Regression that combines beam search with a self-supervised learned heuristic model

Abstract:

Symbolic Regression (SR) aims to discover simple and interpretable mathematical expressions that explain observed data, making it a powerful tool for scientific discovery. In this work, we introduce a conceptually simple SR method that is both sample-efficient with respect to observed data points and self-supervised on large-scale synthetic data. By design, our approach favors parsimony, yielding interpretable and concise expressions. We focus on problems with exact solutions, evaluating our method on datasets containing physical laws and dynamical
equations. Our results demonstrate that combining beam search with a learned heuristic achieves competitive performance compared to existing methods in SRBench. Additionally, our approach effectively handles expressions with constants, a common challenge in the SR field. Finally, we provide a comprehensive scalability analysis across four key dimensions: (i) expression length, (ii) number of variables, (iii) number of domains, and (iv) number of observed data points.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 19041