ELMER: Evolutionary Language Model Enhances Program Evolution via Natural Language

A new method for program evolution called ELMER (Evolutionary Language Model that Explores and Refines) has been developed by researchers, utilizing natural language descriptions to steer mutations. This innovative approach tackles a significant drawback of conventional program evolution: the challenge of controlling mutation movement within behavior space. Often, the size of syntactic edits does not reliably indicate behavioral changes, as minor code modifications can lead to significant behavioral shifts, while extensive rewrites might maintain the same execution trace. ELMER employs a fully fine-tuned Qwen3-8B model to search through natural-language policy descriptions and compile typed programs for execution. This model is trained on three task-specific operations: conditional semantic mutation, translating natural language to domain-specific language (GPTL), and converting GPTL back to natural language. Fine-tuning is conducted with conditional input on mutation strength (low, medium, high) via Direct Preference Optimization (oDPO). In 252 fixed-budget evolutionary searches, oDPO enhanced both behavioral calibration and search efficiency. The peak h-index was recorded using natural language, indicating that language representations can improve evolutionary search. The research paper is accessible on arXiv with the identifier 2608.10196.