Semantic Mutation via LLMs: A Hybrid Approach to Evolutionary Program Synthesis

See solutionsRead the full paper

Semantic Mutation via LLMs: A Hybrid Approach to Evolutionary Program Synthesis

August 4, 2025

Framework diagram

Large language models (LLMs) have recently shown promise for program synthesis, but their use is often hindered by hallucinated outputs, weak adherence to domain-specific languages (DSLs), and reliance on self-evaluation. We propose a hybrid evolutionary framework that integrates LLMs into genetic programming (GP) as untrusted semantic mutation operators.

📄 Paper

Case studyExample

In our framework, the LLM is not treated as a solver or evaluator. Instead, it functions as a mutation operator that proposes candidate programs conditioned on elite individuals, task-specific input-output examples, and execution-based error logs.

Performance chartPanel A

The primary contributions of this work are threefold: semantic mutation via LLMs, a refinement step for synthesis precision, and strong empirical performance on the ARC-AGI benchmark.

🔖 Citation

@article{semantic_mutation_llms,
  title={Semantic Mutation via LLMs: A Hybrid Approach to Evolutionary Program Synthesis},
  author={Anonymous},
  year={2025}
}