T-GADE Evolves LLM Artifacts with Thermodynamic Selection

Summary

The paper introduces T-GADE, a method for evolving structured artifacts produced by large language models, such as a description paired with code. It extends thermodynamical genetic algorithms with LLM-based genetic operators and artifact-level diversity evaluation, while using a common free-energy objective for generational and steady-state updates. Fermi-type occupancy excludes repeated genotypes, whereas Bose-type occupancy allows them. The authors establish an exact one-member removal rule and conditions under which the method recovers the zero-temperature survival rule of Evolution of Heuristics (EoH). On the online bin-packing task used in the EoH paper, training excess measures relative bin-count overhead above a volume lower bound, while transfer excess uses instances with a different bin capacity. Generational Bose-type T-GADE at T=0.003 reduced median training excess by approximately 29%, from 1.152% to 0.815%, across 20 runs per configuration; the reported two-sided Mann-Whitney test gives p=0.042 and Cliff's delta=0.378. When the two highest-ranked final candidates were selected during validation, median transfer excess reached 0.496%, matching EoH. The results support thermodynamical selection and validation-based use of retained artifacts in this task, although the evidence is reported for the studied setting.