ShinkaEvolve - Towards Open-Ended and Sample-Efficient Program Evolution

Explore a groundbreaking open-source framework that leverages large language models for scientific discovery through efficient program evolution in this 47-minute seminar. Learn how ShinkaEvolve addresses critical limitations in current code evolution methods by introducing three key innovations: parent sampling techniques that balance exploration and exploitation, code novelty rejection-sampling for efficient search space exploration, and bandit-based LLM ensemble selection strategies. Discover how this framework achieves state-of-the-art performance with unprecedented sample efficiency, requiring only 150 samples to find new circle packing solutions compared to thousands needed by existing methods. Examine practical applications including the design of high-performing agentic harnesses for AIME mathematical reasoning tasks, improvements to ALE-Bench competitive programming solutions, and the discovery of novel mixture-of-expert load balancing loss functions. Understand how ShinkaEvolve recently supported human programmers in winning the 2025 ICFP Competitive Programming Contest by automatically optimizing SAT solver encodings, demonstrating its real-world impact and broad applicability across diverse computational challenges.