The paper investigates a paradox in Evolution Strategies when tuning large language models: offspring often underperform the parent during training, yet the parent improves. It attributes this to the high-dimensional reward landscape and estimator behavior, where exploration reveals useful directions; aggregating these directions enables local improvements, proposing a noise-curvature optimization paradigm for LLM fine-tuning and high-dimensional problems.
The paper investigates whether coding agents can continually evolve across multiple tasks without catastrophic forgetting, using two harness evolution methods across six directions. It finds that new-task evolution can undermine prior-task performance and offers insights for guiding continual self-evolution, with code available at a referenced URL.
A practical guide teaching ML and optimization via recipes for genetic algorithms, swarms, Monte Carlo methods, and more, with code examples across C++, Python, and JavaScript, and emphasis on testing, visualizing, and applying heuristics and fitness functions.
SwarmEvolve coordinates open-weight LLM agents to generate and improve programs for mathematical and scientific-computing tasks. In tests on 21 problems, it had the highest average score against three open-source frameworks using Qwen3 models, though its code has not yet been released.
SwarmEvolve uses parallel agents that share hidden-state memories to generate and test programs. Across 21 problems, it outscored three open-source baselines on average and reached OpenEvolve’s 60-minute score in a median of 7 minutes, though results varied by task.
GuideEvolve presents a self-evolving agent framework with expert guidance and three regulatory mechanisms (3Ts) to optimize cognition, recover performance via rollback, and terminate adaptively. It outperforms a baseline in accuracy on MGSM and MMLU benchmarks with lower variance and full evolution success, while requiring about half as many optimization iterations and offering improved deployment stability.
SwarmEvolve deploys hundreds of latent-space AI agents that collaborate in swarms to rapidly discover and optimize solutions for mathematical and scientific computing tasks, without text decoding. The framework uses latent-state sharing, a novelty filter, and natural idea crossover to outperform prior methods in both speed and accuracy.
SkillNEO introduces non-elitist evolutionary optimization for agent skills, using execution feedback to iteratively refine versions. It preserves useful partial progress, revisits earlier work with new evidence, and demonstrates performance gains across multiple models and domains.
A study shows EnsembleML uses an ensemble of four regressors trained on historical Pareto fronts to predict changes in dynamic multi-objective optimization, reinitializing the population with Gaussian perturbation and mutation. Evaluated on the CEC 2018 benchmark against eight methods, it often performs best on MIGD and MHVD, though results vary by problem and change severity, suggesting potential but limited real-world applicability.
The paper introduces an agent-data co-evolution framework, \method, linking task design, rollout learning, and fine-tuning to adapt MCP agents to evolving student capabilities; iterative loops of evaluation and data evolution yield consistent performance gains across multiple benchmarks.