Technical Summary: ARES-LSHADE

Problem Statement

The paper addresses the GECCO 2026 competition challenge: designing an optimization algorithm with explicit Large Language Model (LLM) involvement in the design process, evaluated on the Generalized Numerical Benchmark Generator (GNBG). The GNBG suite consists of 24 continuous minimization problems constructed as the minimum of several rotated, transformed quadratic components. Key constraints include:

• Blackbox Treatment: The algorithm must rely solely on function evaluations; it cannot access the internal parameters defining the function's structure (e.g., component minimum positions).
• LLM Involvement: An LLM must be used during the algorithm's design phase.
• Evaluation: Algorithms are tested on 31 independent runs per function with specific function-evaluation budgets (500,000 for functions f1–f15; 1,000,000 for f16–f24). A "win" is defined as achieving an absolute gap below $10^{-8}$ from the true optimum.

Methodology

ARES-LSHADE (Autoresearch-Enhanced LSHADE) is a memetic differential evolution variant that builds upon the 2025 winner, LLM-LSHADE. The architecture consists of three primary components:

1. Inherited Framework

The algorithm retains the core LSHADE framework, including:

• Initialization and Crossover: Adaptive strategies (uniform, prime-spiral, Sobol-plus-cluster) and a bracket-adaptive crossover operator designed by LLM-LSHADE in 2025.
• Population Dynamics: Linear population size reduction (LPSR) from ~180 to 4.

2. Autonomous Research Loop (Mutation Operator)

A novel component is the mutation operator, generated by an autonomous LLM-driven research loop.

• Architecture: The loop ran approximately 30 design experiments over several weeks. In each iteration, the LLM (Claude) received the current operator source, a CSV of benchmark results, and an experiment log. It proposed a new Python mutation operator, which was validated and tested.
• Constraints: The loop's edit surface was restricted to the mutation operator only. The observation space included fitness traces, landscape descriptors ($\lambda, \omega$, rotation), and the experiment log, but explicitly excluded the Component_MinimumPosition metadata found in GNBG problem files.
• Outcome: The loop converged to a stable plateau of 16–17 wins out of 24. The final submitted operator combines current-to-pbest/1 with an archive, a 20% "scout" subpopulation for diversity, and a 10% CMA-ES branch for rotated single-component functions.

3. Memetic Polish Phase

A multi-start local search phase runs in the final 5% of the function-evaluation budget.

• Mechanism: It executes L-BFGS-B local optimizations from eight starting points derived strictly from the Evolutionary Algorithm's (EA) discoveries and public search bounds:
  1. The EA's best-found point.
  2. Four perturbations of the best point (at 2%, 5%, 10%, and 20% of the search domain width).
  3. Three uniform random restarts.
• Blackbox Compliance: Crucially, the submitted version does not seed the local search using the Component_MinimumPosition metadata. The authors note that an earlier variant using this metadata achieved 24/24 wins but was discarded as it violated the blackbox rule by effectively "reading" the solution from the problem file rather than discovering it through evaluation.

Key Results

On the official 31-run-per-function evaluation:

• Total Wins: 510 out of 744 possible wins.
• Machine Precision: 15 of the 24 functions were solved to machine precision ($<10^{-8}$ gap) on all 31 runs.
• Failure Analysis: Six functions (f6, f13, f14, f15, f21, f24) consistently failed to meet the threshold. These failures were not random but exhibited distinct signatures:
  • Deterministic Near-Miss (f21): A constant gap of 5.0 with zero variance, indicating the algorithm converged to a specific local minimum.
  • Tight Near-Miss (f6, f15): Low-variance gaps (0.36 and 5.4), suggesting the algorithm found a basin but not the global one.
  • High-Variance Basin Search (f13, f14, f24): Large gaps with high variance, indicating the search landed in different deceptive basins across runs.
• Correlation: The six failing functions were independently identified as the "hardest" by the autoresearch loop during its design phase, suggesting the loop correctly diagnosed the difficulty limits of the architecture.

Significance and Methodological Claims

The paper positions its contribution not merely as a high-performing algorithm, but as a methodological study on LLM-driven algorithm design:

1. Plateau Diagnosis: The authors argue that the six failing functions represent a characteristic plateau for population-based memetic algorithms without structural priors on the GNBG benchmark. The failures are diagnostically informative, revealing specific compositional structures (e.g., plateaus, small global basins) that are difficult to overcome under strict blackbox constraints.
2. Edit Surface and Observation Space: The work highlights that the "bottleneck" in LLM-driven design is often the choice of the edit surface (what the LLM can change) and observation space (what the LLM can see). The loop's inability to break the 16–17 win plateau is attributed to its restriction to operator-only edits and the lack of aggregate cross-function failure data in its prompt.
3. Integrity vs. Capability: The paper documents a critical tension between LLM capabilities and benchmark integrity. The authors demonstrate that when an LLM is given access to benchmark metadata (like component positions), it can trivially solve the problem, producing scores that reflect the benchmark's internal structure rather than the algorithm's search efficacy. They advocate for explicit "data-access disclosures" in future LLM-driven research to prevent such "oracle exploitation."

In conclusion, ARES-LSHADE achieves a strong performance (510/744 wins) under strict blackbox rules, but the paper's primary value lies in its transparent documentation of the design process, the identification of algorithmic limits on specific benchmark structures, and the ethical handling of metadata access.