Large language models (LLMs) are rapidly transforming how we create, reason about, and automate AI techniques and algorithmic discoveries. Instead of merely tuning hyperparameters or selecting among existing methods, LLMs enable fully automated algorithm design (AAD), the discovery of new architectures, and even end-to-end pipeline construction, effectively closing the loop between problem description, ideation, implementation, and evaluation in natural language. In parallel, evolutionary computation (EC) offers a powerful toolbox for exploring vast, high-dimensional and multi-modal search spaces, evolving not only solutions but also operators, representations, and full algorithmic workflows, often under noisy, constrained, or partially known conditions. EC has a long tradition of open-ended search, robustness, and creativity in optimization and design, making it an ideal counterpart to LLM-driven generation and reasoning. Together, these developments invite a rethinking of how we design, adapt, and explain optimization and learning systems—moving from hand-crafted pipelines toward hybrid, self-improving AI design ecosystems. What connects these two together?

One answer is evolutionary search heuristics powered by LLMs (“LLMs with EC”), where the model repeatedly generates and refines candidate algorithms, operators, or configurations, while selected evolutionary principles - populations, selection, crossover, and mutation- guide their improvement. Some of the recent frameworks, such as LLaMEA, FunSearch, AlphaEvolve, EASE, and EOH, may fully/partially illustrate this setup: an LLM proposes a population of candidate heuristics (often in code), which are evaluated (on benchmark tasks), and high-performing candidates are recombined and mutated in subsequent rounds. A shared knowledge base of evaluated solutions and reusable components accumulates over time, while feedback signals steer both the evolutionary search and the orchestration of modules around the LLM. In this way, EC provides structured exploration and selection pressure, and the LLM acts as a versatile generator and refiner inside a closed, iterative loop. This hybridization overturns the conventional paradigm that ECs use, and in turn, sometimes yields high-performing and novel EC systems. Of course, there may be more views on LLMs with EC, spanning from benchmarking of discovered algorithms, evolutionary prompt engineering, optimization of LLM-based frameworks and architectures, and more.

Another approach is to use LLMs for EC. LLMs may help researchers select feasible candidates from the pool of algorithms based on user-specified goals, a description of optimization tasks, and a characterization of available features, and provide a basic description of the methods or propose novel hybrid methods. Furthermore, the models can help identify and describe distinct components suitable for adaptive enhancement or hybridization, and provide pseudo-code, implementation, and reasoning for the proposed methodology.

Ultimately, by combining both answers, LLMs have the potential to transform automated metaheuristic design as part of AAD and configuration by generating EC module codes, and iteratively improving the initially designed solutions or algorithm templates (with or without performance or other data-driven feedback).

This workshop aims to encourage innovative approaches that leverage the strengths of LLMs and EC techniques, thus enabling the creation of more adaptive, efficient, and scalable algorithms by integrating evolutionary mechanisms with advanced LLM capabilities. Thanks to the collaborative platform for researchers and practitioners, the workshop may inspire novel research directions that could reshape AI, specifically LLMs, and optimization fields through this hybridization and achieve a better understanding and explanation of how these two seemingly disparate fields are related and how knowledge of their functions and operations can be leveraged.

It includes (but is not restricted to the following topics):

• LLM-based iterative frameworks (with evolutionary principles)
• Benchmarking of generted metaheuristics
• Evolutionary Prompt Engineering
• Optimisation of LLM Architectures
• LLM-Guided Evolutionary Algorithms
• How can an EA using an LLM evolve different of units of evolution, e.g. code, strings, images, multi-modal candidates?
• How can an EA using an LLM solve prompt composition or other LLM development and use challenges?
• How can an EA using an LLM integrate design explorations related to cooperation, modularity, reuse, or competition?
• How can an EA using an LLM model biology?
• How can an EA using an LLM intrinsically, or with guidance, support open-ended evolution?
• What new variants hybridizing EC and/or another search heuristic are possible and in what respects are they advantageous?
• What are new ways of using LLMs for evolutionary operators, e.g. new ways of generating variation through LLMs, as with LMX or ELM, or new ways of using LLMs for selection, as with e.g. Quality-Diversity through AI Feedback)
• How well does an EA using an LLM scale with population size and problem complexity?
• What is the most accurate computational complexity of an EA using an LLM?
• What makes good EA plus LLM benchmark?
• LLMs for (automated) generation of EC.
• Understanding, fine-tuning, and adaptation of Large Language Models for EC. How large do LLMs need to be? Are there benefits for using larger/smaller ones? Ones trained on different datasets or in different ways?
• Implementing/generating methodology for population dynamics analysis, population diversity measures, control, and analysis and visualization.
• Generating rules for EC (boundary and constraints handling strategies).
• The performance improvement, testing, and efficiency of the improved algorithms.
• Reasoning for component-wise analysis of algorithms.
• Connection of LLM and other ML techniques for EC (Reinforcement learning, AutoML)
• Generation and reasoning for parallel approaches for EC algorithms.
• Benchmarking and Comparative Studies of LLM-generated algorithms.
• Applications of LLM and EC (not limited to):
  • constrained optimization
  • multi-objective optimization
  • expensive and surrogate assisted optimization
  • dynamic and uncertain optimization
  • large-scale optimization
  • combinatorial/discrete optimization