ARES-LSHADE: Autonomous Research Enhances Evolutionary Algorithm for GNBG Benchmark
ARES-LSHADE, a variant of memetic differential evolution, has been entered into the GECCO 2026 competition focused on LLM-created evolutionary algorithms for the Generalized Numerical Benchmark Generator (GNBG). This new entry builds upon the success of LLM-LSHADE, the 2025 champion, by incorporating two innovative features: a scout-augmented mutation operator that integrates adaptive CMA-ES, developed from around thirty LLM-guided design experiments, and a multi-start L-BFGS-B polishing phase that adheres to strict blackbox protocols. In official assessments, with 31 evaluations per function and set budgets, ARES-LSHADE secured 510 victories out of 744 (with per-function gaps under 1e-8), achieving machine precision on 18 out of 24 functions. The other six functions exhibited plateau signatures aligned with GNBG's compositional framework, as identified by the autonomous research loop.
Key facts
- ARES-LSHADE is a memetic differential-evolution variant submitted to GECCO 2026 competition.
- It builds on the LLM-LSHADE 2025 winner.
- Two new components: scout-augmented mutation with adaptive CMA-ES integration and multi-start L-BFGS-B polish.
- Autonomous research loop involved approximately thirty LLM-driven design experiments.
- Achieved 510 of 744 wins on official 31-run-per-function evaluation.
- Reached machine precision on 18 of 24 functions.
- Remaining six functions showed plateau signatures consistent with GNBG's compositional structure.
- Competition is on LLM-designed evolutionary algorithms for GNBG benchmark.
Entities
Institutions
- GECCO
- LLM-LSHADE