ELMER: Evolutionary Language Model that Explores and Refines

2026-08-10

AI Summary

This work addresses the limitations of existing program evolution methods, which struggle to effectively control the magnitude of behavioral variation and rely on syntactic edit distances that poorly capture semantic change. To overcome these challenges, the authors propose a large language model–based program evolution framework that introduces natural language as a guided, executable intermediate representation. By fine-tuning Qwen3-8B with task-conditioned supervision and orthogonal Direct Preference Optimization (oDPO), the framework enables conditional semantic variation and bidirectional compilation between natural language and a domain-specific language (GPTL). Evaluated under a fixed budget of 252 search iterations, the approach significantly improves behavioral alignment and search efficiency, with natural language representations achieving optimal fitness retention and enabling synergistic optimization of behavioral displacement and semantic fidelity.

Abstract

Program evolution can measure whether a mutation helped, but it rarely controls how far the mutation moves in behavior space. Syntactic edit size is an unreliable proxy: a small code change can alter nearly every action, while a larger rewrite can preserve the same execution trace. We introduce an Evolutionary Language Model that searches over natural-language policy descriptions and compiles typed programs for execution. A fully fine-tuned Qwen3-8B model learns three task-conditioned operations: conditional semantic mutation, natural language to domain-specific language (GPTL) compilation, and GPTL to natural language translation. The model is fine-tuned with conditional input on the mutation strength (low, medium, high) using Direct Preference Optimization (oDPO). Across 252 fixed-budget evolutionary searches, oDPO improves both behavioral calibration and finite-budget search efficiency. Natural-language attains the highest observed held-out fitness. Our analysis shows that the condition input (mutation strength) systematically changes semantic edit composition and that language mutations preserve more parent fitness at matched small-to-moderate behavioral displacement. These results show that language can serve as a steerable, execution-grounded search representation over executable program space.

Problem

• program evolution
• behavioral displacement
• semantic mutation
• language-guided search
• mutation strength

Innovation

• Evolutionary Language Model
• Semantic Mutation
• Natural Language Programming
• Behavioral Calibration
• Direct Preference Optimization