Deliberate Evolution: Agentic Reasoning for Sample-Efficient Symbolic Regression with LLMs

Researchers have introduced new methods for symbolic regression, a technique used to find symbolic expressions that describe datasets. The methods enable the description of related phenomena with varying sets of parameters.

Symbolic regression (SR) is a powerful paradigm for scientific discovery due to its inherent interpretability. Recent advances have expanded SR to describe related phenomena using a single expression with varying sets of parameters. One research team, publishing on arXiv^[1], introduced partial parameter sharing, allowing for multiple categorical variables and intermediate levels of parameter sharing. This enables the separation of universal effects, category-specific trends, and category interactions. The method achieved similar fit quality with significantly fewer parameters in an astrophysics dataset. Another team, also on arXiv^[2], proposed Deliberate Evolution, an agentic framework for sample-efficient symbolic regression with Large Language Models (LLMs). Deliberate Evolution decouples symbolic generation from search control and guides LLM proposals with adaptive operators and reflective memory. Experiments showed that Deliberate Evolution outperformed LLM-based SR baselines across diverse scientific domains, using only 40% of the standard sample budget^[2].