T-GADE
An LLM evolution framework that picks survivors by minimizing a 'free energy' so a temperature setting trades off quality against diversity.
The loop
- Proposes
Structured artifacts (here, bin packing scoring heuristics) from LLM crossover, mutation and a self-repair step
- Judged by
Evaluation on online bin packing, measured as excess bins over the volume lower bound
- Steers
Survival selection minimizes free energy F_T = <E> - T*H, with temperature T controlling the quality-diversity balance
Headline result
In a preregistered seed-paired comparison at 400 total LLM calls on online bin packing, the HoxLM configuration won 9 of 10 runs against EoH, with a median sealed excess of 0.836% against 3.912%.
What this does not prove
Preprint (arXiv v2, 14 Oct 2026), one task, 10 seed pairs, and an EoH-aligned implementation. Not independently replicated.
How it generalizes the harness
A physics-inspired selection rule over a population steers the search, and the evolved scoring rules are readable and run without the model.
Classification
| What it proposes | Heuristic |
|---|---|
| Generative model | Pretrained LLM, frozen (also used as a self-repair model, HoxLM) |
| Algorithmic authority | Bin packing evaluation and free-energy survival selection |
| Feedback | Score |
| What gets updated | Population of candidates |
| When it runs | Offline, at design time |
| Validity | Empirical (measured) |
| Evidence | Experiments reported |
| Lineage | Derivative |