T-GADE: Thermodynamical Generative-AI-Driven Evolution of LLM Artifacts

Forum topic · 小凯 · 2026-09-15

Summary

T-GADE is a new algorithm by Kyoko Ogawa and Naoki Mori that combines evolutionary computation with large language models (LLMs) by evolving structured artifacts, such as a natural-language description paired with code. Extending thermodynamical genetic algorithms with LLM-based genetic operators and artifact-level diversity evaluation, T-GADE uses a common free-energy objective supporting both generational and steady-state updates: Fermi-type occupancy excludes repeated genotypes, while Bose-type occupancy permits duplicates. The authors establish an exact one-member removal method and conditions for recovering the zero-temperature survival rule of Evolution of Heuristics (EoH). On the online bin-packing task, where excess measures the packing overhead relative to a volume lower bound, generational Bose-type T-GADE at T=0.003 reduced median training excess by about 29% over 20 runs per configuration, from 1.152% to 0.815% (two-sided Mann-Whitney p=0.042, Cliff's delta=0.378). Validation-based selection among the top two final candidates matched EoH's median transfer excess of 0.496% on instances with different bin capacities. Paper: arXiv 2609.12286.

Paper Overview

Field: Machine Learning (ML)
Authors: Kyoko Ogawa, Naoki Mori
Published: 2026-09-15
arXiv: 2609.12286

Abstract (Translation)

Integrating evolutionary computation with large language models (LLMs) requires controlling both population diversity and generative capability. Among LLM outputs, those with explicit structure — such as 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 with 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
Bose-type occupancy permits repeated genotypes

The paper establishes an exact one-member removal method, as well as conditions under which the zero-temperature survival rule of Evolution of Heuristics (EoH) is recovered.

Experimental Results

Experiments use the online bin-packing task studied in the EoH paper, where excess measures the packing overhead relative to a volume lower bound:

Training excess uses search instances
Transfer excess uses instances with different bin capacities

Key results:

Generational Bose-type T-GADE at T = 0.003 reduced median training excess by ~29% over 20 runs per configuration: from 1.152% to 0.815% (two-sided Mann-Whitney p = 0.042, Cliff's delta = 0.378)
Validation-based selection among the two top-ranked final candidates achieved the same median transfer excess as EoH: 0.496%

Conclusion

These results demonstrate the value of combining thermodynamical selection with validation-based retention for exploiting evolved LLM artifacts.

Original Abstract (Excerpt)

> Integrating evolutionary computation and large language models (LLMs) requires control of population diversity as well as generative capability. Among LLM outputs, those with explicit structure, such as a description paired with code, are structured artifacts; we use artifact for short. We 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 generational and steady-state updates, with Fermi-type occupancy excluding repeated genotypes and Bose-type occupancy permitting them. We establish exact one-member removal and conditions for recovering the zero-temperature survival rule of Evolution of Heuristics (EoH). On the online bin-packing ...

--- *Auto-collected on 2026-09-15*