LLMs beat GP baselines with zero-shot parent selection operators

A new study from researchers at Victoria University of Wellington and Michigan State University explores whether large language models (LLMs) can generate effective parent-selection operators for genetic programming (GP) in symbolic regression—without any iterative feedback or reference operators. The zero-shot approach asks each LLM to write a simple Python function for selecting parents during evolution, then drops it into a standard GP framework. Eight models were tested (including Claude, Gemini, GPT-4, and Kimi), each generating 10 independent operators evaluated on 12 regression benchmarks from OpenML.

The results show that Claude Sonnet 4.6 and Gemini 3.1 Pro deliver consistently strong training and held-out test R² scores. The single best operator came from Kimi K2.5, which beat the popular automatic lexicase and tournament selection baselines. Many LLM-generated operators use semantic information (e.g., prediction error or diversity) to guide selection, indicating that LLMs can infer non-trivial search heuristics from the task description alone. The study also found that public leaderboard rankings like Humanity's Last Exam and SWE-bench Verified correlate well with training performance but less clearly with generalization, suggesting caution when using generic benchmarks to predict evolutionary algorithm effectiveness.