gepa-ai.github.io
|
ksl
|
|
A UC Berkeley team led by Matei Zaharia released optimize_anything, a declarative API that extends their GEPA prompt optimizer to handle code, agent architectures, CUDA kernels, and system configurations. You provide a seed artifact and an evaluator function, and the framework iteratively refines it using LLM reasoning over diagnostic feedback. The benchmark results are hard to ignore — ARC-AGI accuracy jumped from 32.5% to 89.5% with Gemini 3 Flash, and it outperformed AlphaEvolve on circle packing. Infrastructure cost reductions hit 40% in cloud optimization tests. The tool sits in a growing category alongside Google’s AlphaEvolve and similar evolutionary search frameworks, but the pip-installable API and backend-agnostic design lower the barrier enough that applied teams can actually use it without rebuilding their stack.
