

Neuroevolution with Geometric & ETH Zürich ASL
What happens when AI can propose software improvements faster than we can review them? Join us at ETH Zürich for two technical talks: one on using coding agents to deliver measurable speedups, and one on how agents can learn abstractions that make it easier to search for better solutions.
Automating software optimization · Max Buckley
When generating patches is no longer the bottleneck, how do you decide what’s worth merging? Max will show how to coordinate coding agents in parallel, use fast evaluations to decide which changes are worth keeping, and have an agent from a different vendor challenge each patch. He’ll walk through three case studies: GPU-accelerated LightGBM, the SCIP mixed-integer solver, and LLM inference kernels. You’ll leave with a workflow for finding and verifying speedups in your own codebase.
Learning Hierarchical Languages and World Models · Boris Albar
World models let agents reason about what might happen before they act. Boris will present a compiler-inspired approach in which agents learn layers of abstraction grounded in formal models they can revise. He’ll explore how agents can choose which level to search, rule out equivalent or impossible candidates, and judge whether an abstraction is useful by how much reasoning it saves, whether proofs can be reused, and how well it transfers to new problems.
Kevin O’Sullivan will also give a short introduction to ETH Zürich’s Agentic Systems Lab.