Motivation
Large language models (LLMs) are transforming AutoML—not just optimizing hyperparameters, but generating entirely new algorithms. This tutorial provides a crash course on this fast-evolving landscape. We cover foundational systems like LLaMEA, EASE, and AlphaEvolve, and guide participants through safe, reproducible experimentation using open-source tools and benchmarks.
LLM-based algorithm design is one of the fastest-moving frontiers of AutoML. In just two years, we’ve gone from simple prompt engineering to complex systems like AlphaEvolve that can co-evolve entire codebases. Tools like FunSearch are rewriting proofs, while ReEvo and LLaMEA evolve optimization heuristics on the fly.
But despite the excitement, most researchers still struggle with practical questions:
- Which frameworks are usable today?
- How do I use these automated algorithm discovery tools?
- What is required for these tools to work, and on what problems can I apply them?
- How do I evaluate these tools and their generated outcomes?
This tutorial closes that gap. We’ll distill the core ideas behind each approach, walk through runnable code, and highlight challenges in safety, reproducibility, and benchmarking. It’s not just a demo—we aim to equip you with conceptual tools and practical workflows to contribute to this new generation of AutoML.
Attendees will gain hands-on skills in designing, running, and evaluating LLM-driven algorithm discovery pipelines using Colab notebooks and lightweight evolutionary frameworks.
Program Outline
Program Outline (90 minutes)
- 10 min: Introduction & landscape – why algorithm discovery matters for AutoML
- 15 min: Core paradigms: FunSearch, LLaMEA, ReEvo, AlphaEvolve, EASE
- 20 min: Deep dive into LLaMEA, LLaMEA-HPO & LLaMEA-BO
- 15 min: EASE: frontend/backend architecture + live examples
- 10 min: Evaluation & safety: BLADE, benchmarks, risk mitigation
- 10 min: Frontiers & open challenges (bias, compute cost, integration with pipelines)
- 10 min: Open Q&A
