Researchers unveil optimize_anything: a universal LLM API for text optimization

A team of 15 researchers (including Dan Klein, Ion Stoica, Joseph Gonzalez, and Matei Zaharia) has open-sourced optimize_anything, a universal API that frames optimization as improving a text artifact evaluated by a scoring function. Rather than building separate tools for agent architectures, cloud scheduling, CUDA kernels, or circle packing, a single LLM-based search system tackles all six domains. Key results include nearly tripling Gemini Flash's ARC-AGI accuracy from 32.5% to 89.5%, discovering scheduling algorithms that cut cloud costs by 40%, and generating CUDA kernels where 87% match or beat PyTorch's performance. The system even outperforms AlphaEvolve's reported circle packing solution for n=26.

The researchers found that providing actionable side information (e.g., error messages or partial solutions) leads to faster convergence and higher final scores compared to score-only feedback. Additionally, multi-task search—training on related problems simultaneously—consistently outperforms independent optimization given the same per-problem budget, with benefits scaling as more related tasks are added. The work demonstrates for the first time that text optimization with LLM-based search is a general-purpose problem-solving paradigm, unifying tasks that previously required domain-specific algorithms. optimize_anything is released as part of the GEPA project with support for multiple backends, making it accessible for researchers and practitioners to adapt to their own optimization problems.