The research presented T-GADE, a method for evolving LLM artifacts using a thermodynamical genetic algorithm. The system focused on structured artifacts, such as descriptions paired with code, and utilized LLMs to drive the evolutionary process through genetic operators. A common free-energy objective, employing Fermi-type and Bose-type occupancies, guided generational and steady-state updates, aiming to minimize repeated genotypes while allowing for diversity.
Experiments were conducted on the online bin-packing task, a benchmark previously used in the Evolution of Heuristics (EoH) paper. T-GADE, operating at a temperature of 0.003, achieved a reduction of approximately 29% in median training excess, decreasing it from 1.152% to 0.815% across 20 runs. Statistical analysis, using a two-sided Mann-Whitney test with a p-value of 0.042 and a Cliff's delta of 0.378, confirmed the significance of this reduction.
Furthermore, validation was performed by selecting the two highest-ranked final candidates. The median transfer excess achieved with this validation approach matched the performance of EoH, reaching 0.496%. This demonstrated the utility of a validation-based approach for retaining and utilizing evolved artifacts.
The system addressed the control of population diversity inherent in integrating evolutionary computation with LLMs. The use of thermodynamic selection, guided by free-energy objectives, provided a mechanism for efficient artifact evolution. The results suggest a pathway for optimizing LLM artifact generation through a combination of generative AI and evolutionary techniques.
Source: https://arxiv.org/abs/2609.12286
