Overview
Field: Machine Learning Authors: Kyoko Ogawa, Naoki Mori Published: 2026-09-15 arXiv: 2609.12286
Summary
Integrating evolutionary computation with large language models (LLMs) requires controlling population diversity as well as generative capability. LLM outputs with explicit structure — for example, a description paired with code — are called structured artifacts (or simply artifacts). The authors propose T-GADE, which evolves these artifacts by extending thermodynamical genetic algorithms through LLM-based genetic operators and artifact-level diversity evaluation.
A common free-energy objective supports both generational and steady-state updates: Fermi-type occupancy excludes repeated genotypes, while Bose-type occupancy permits them. The paper establishes exact one-member removal and conditions for recovering the zero-temperature survival rule of Evolution of Heuristics (EoH).
Experimental Results
Experiments use the online bin-packing task studied in the EoH paper, where excess measures the overhead in number of bins relative to a volume-based lower bound (training excess uses search instances; transfer excess uses instances with different bin capacities).
- The generational Bose-type T-GADE with T=0.003 reduced the median training excess by ~29%, from 1.152% to 0.815%, over 20 runs per configuration (two-sided Mann-Whitney p=0.042, Cliff's delta=0.378).
- Validation-based selection among the two highest-ranked final candidates achieved the same median transfer excess as EoH: 0.496%.
Conclusion
These results demonstrate the value of combining thermodynamic selection with validation-based exploitation of retained artifacts for LLM-driven evolutionary search.
--- *Auto-collected on 2026-09-15*