LLMfwEC Workshop at GECCO 2027
Pending acceptance
This event has been proposed and is awaiting acceptance by the conference organisers. Details may change.
We have proposed the 3rd Workshop on Large Language Models for and with Evolutionary Computation (LLMfwEC) at GECCO 2027, 12–16 July 2027 in Kraków, Poland. This is the third edition of the workshop at GECCO.
Large Language Models for and with Evolutionary Computation
Large language models (LLMs) are changing how algorithms are conceived, written and evaluated. Frameworks such as FunSearch, AlphaEvolve, EoH, EASE and LLaMEA pair an LLM with evolutionary principles, and they now generate and refine candidate algorithms, operators, prompts and whole pipelines, mostly as code. Population, selection and variation provide structured exploration and selection pressure, and the LLM acts as a flexible generator and refiner inside a closed loop of generation, evaluation and feedback. In the opposite direction, LLMs increasingly support evolutionary computation (EC) researchers and practitioners by helping to choose, hybridise, configure, implement and explain metaheuristics.
LLMfwEC brings these two views together: EC with LLMs (LLMs as engines of evolutionary search) and LLMs for EC (LLMs as assistants for designing, analysing and applying EC). After the first wave of “it can be done” results, the community now has to answer harder questions. Are the discovered algorithms genuinely new and better, and under which benchmarks and budgets? What did the LLM actually contribute, and can we explain why an evolved algorithm works? What does it cost in compute and money, and how do smaller models and ensembles compare with large ones? Can the results be reproduced and trusted? This edition puts these questions of understanding, benchmarking and trust at the centre, alongside new methods and applications.
The workshop is a meeting point for researchers and practitioners from EC, machine learning, program synthesis and explainable AI.
Topics of Interest
Topics include, but are not limited to:
- LLM-driven evolutionary frameworks for automated algorithm design, evolving code, text, images and multimodal candidates
- LLMs as evolutionary operators: variation, crossover and mutation, selection, quality-diversity with AI feedback, open-ended evolution
- Benchmarking and comparative studies of LLM-generated algorithms: benchmark design, data contamination, statistical comparison, reproducibility
- Explainability and analysis of discovered algorithms: component-wise analysis, ablations, search-trajectory and population-dynamics analysis, attributing performance to the LLM
- Evolutionary prompt engineering and the optimisation of LLM-based agents, pipelines and architectures
- Models, cost and scaling: small versus large models, fine-tuning, orchestration of model ensembles, token, energy and budget efficiency, scaling with population size and problem complexity
- Reliability of generated code: verification, robustness, failure modes and safety in evaluation-in-the-loop systems
- Cooperation, modularity, reuse and competition in LLM-based evolutionary systems; modelling biology
- LLMs as assistants for EC practitioners: algorithm selection, hybridisation, configuration, constraint and boundary handling, parallel implementations, diversity measures and visualisation
- Combining LLMs with other learning and search techniques such as reinforcement learning and AutoML
- Applications: constrained, multi-objective, expensive and surrogate-assisted, dynamic and uncertain, large-scale and combinatorial optimisation, and real-world design problems
Submissions
We plan to accept two types of submissions:
- Full papers (8 pages + references), presented as 10-minute pitches
- Extended abstracts (up to 4 pages), presented as 5-minute pitches
Accepted papers will appear in the GECCO 2027 Companion volume in the ACM Digital Library. Submissions follow the ACM and GECCO policy on generative AI.
An Interactive Session
- Invited keynote on LLM-driven discovery, open-endedness or program synthesis.
- Panel discussion with the audience on benchmarking LLM-designed algorithms, explaining what the LLM contributed, and models, cost and scaling, steered by audience questions and live polls.
- Shared roadmap board where participants collect and vote on the most pressing open problems.
Key Dates
To be announced once the workshop is accepted. GECCO 2027 takes place 12–16 July 2027 in Kraków, Poland.
Previous Editions
Organisers
- Niki van Stein, Leiden University, the Netherlands (contact)
- Roman Senkerik, Tomas Bata University in Zlín, Czech Republic
- Michal Pluhacek, AGH University of Krakow, Poland
- Erik Hemberg, MIT CSAIL, USA
- Anna V. Kononova, Leiden University, the Netherlands
- Pier Luca Lanzi, Politecnico di Milano, Italy
- Joel Lehman, independent researcher
- Tome Eftimov, Jožef Stefan Institute, Slovenia
Acknowledgements
This event is made possible by the IEEE CIS Task Force on Generative AI for Evolutionary Computation (GenAI4EC).
Organizers:
Associate Professor of Explainable AI
Assistant Professor of Efficient Heuristic Optimization
