Jul 2026· Proceedings of the Genetic and Evolutionary Computation Conference Companion· pp. 13-14· 0 citations· 3 references
TL;DR
It is argued that model diversity is a usable, low-cost resource for LLM-driven algorithm design.
Abstract
In this entry to the GECCO 2026 Competition on LLM-Designed Metaheuristics, three large language models (GPT-5.4 Thinking, Gemini 3.1 Pro, Claude Opus 4.6) were independently prompted to design a metaheuristic for the GNBG benchmark (24 problems, hence the /24 scoring scale). Each reached a per-model ceiling (GPT 19.82, Claude 21.08, Gemini 23.29) and could not improve further on its own. Handing the best design (Gemini's) to the second-best model (Claude) for refinement produced, within four iterations, EO-BIPOP-CMA-CDR, which ranks first in the full 24-algorithm field at 23.30/24 and outperforms the source design head-to-head by three score points. A single cross-model handoff yielded more progress than five further self-refinement iterations by the source model. We argue that model diversity is a usable, low-cost resource for LLM-driven algorithm design.
It is found that neutral framing of behavioural features, reporting the feature value without prescriptive advice, consistently outperforms the prescriptive variants, and that prescriptive feedback steers the median feature value in the advised direction in only 37% of cases despite empirically grounded advice.
It is found that none of the LLaMEA generated algorithms outperformed KMeans++, and prompting strategy affects the structural diversity of generated algorithms, with exemplar-injection producing the broadest range of algorithm families.
This paper compares six different strategies identified in a recent survey, categorizing them as either global strategies (GS), which treat the decision vector as a unified type, or decomposition-based strategies (DS), which split the problem into continuous and discrete subproblems.
Thomas Lang, Denis Pallez· GECCO Companion· 0 citations
This study introduces a standardized framework based on a Normalized Positional Diversity Index (D*) to quantify optimizer behaviour and demonstrates that D* is a geometric generalization of existing measures, by replacing stochastic, path-dependent historical maximums with a fixed global upper bound anchored to the search space geometry.
Metaheuristics require sustained global search without sacrificing local refinement, yet many variable-structure methods change operators through one-way iteration schedules. We introduce the Weather State Ants Optimizer (WSAO), in which a discrete-time Markov chain recurrently selects one of three population updates. Sunny, cloudy, and rainy states correspond to global exploration, movement toward nests, and local refinement, respectively. An archive-based mechanism also maintains several spatially separated nests as concurrent search centers. Thirty independent runs compared WSAO with 11 algorithms on 29 CEC2017 and 12 CEC2022 functions. WSAO achieved the lowest Friedman mean rank on both suites, at 2.48 and 2.33. Across five constrained design cases, it joined the leading group by mean objective value on four cases and ranked second on pressure-vessel design. Targeted CEC2022 controls showed that no alternative transition matrix dominated the baseline. Eliminating the trial perturbation worsened every selected function, whereas the contribution of multiple nests depended on the landscape structure. The combined evidence supports recurrent state-controlled search as a competitive framework for continuous numerical and constrained optimization.
The results suggest that recent progress in language models and tool use may already be sufficient to support practical automated metaheuristic design, and that recent progress in language models and tool use may already be sufficient to support practical automated metaheuristic design.
Jan Iłowski, Marcin Małek, Wojciech Achtelik et al.· GECCO Companion· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
