Morning
8:30 AM – 12:30 PMIn this talk, I'll share the story of ShinkaEvolve, our open-source framework for sample-efficient program evolution, and reflect on how it fits into a broader research trajectory toward LLM- and agent-driven scientific discovery. I'll start with the motivation: recent progress in scaling inference-time compute has made evolutionary agentic harnesses a surprisingly powerful tool for discovery, but existing systems are sample-inefficient and largely closed-source. I'll then walk through Shinka's core recipe: parent program sampling that balances exploration and exploitation, novelty-based rejection sampling, and bandit-driven LLM ensemble selection. These ingredients let Shinka discover a new state-of-the-art circle packing with just 150 samples, design strong mathematical reasoning harnesses, improve competitive programming solutions, and, most recently, help team Unagi win the 2025 ICFP Programming Contest by optimizing SAT encodings. In the second half, I'll focus on what's changed since Shinka's initial release: substantial throughput optimizations, a cost-aware model selection mechanism that makes ensembles economically sane, and a new CLI that slots Shinka directly into general-purpose coding agents like Claude Code and Codex, turning program evolution into something you can smoothly invoke during everyday development.
Rob is a Research Scientist and founding member at Sakana AI. He is also a final-year PhD student working on Evolutionary Meta-Learning at the Technical University Berlin. Previously, he completed an MSc in Computing at Imperial College London, a Data Science MSc at Universitat Pompeu Fabra, and an Economics undergraduate at University of Cologne. He worked at Google DeepMind with the Tokyo team as a full-time student researcher and interned at Legacy DeepMind (Discovery team) & Accenture. He maintains a set of open source tools including evosax (JAX-based Evolution Strategies) and gymnax (JAX-based Reinforcement Learning Environments).