optimize_anything unifies six optimization domains under one LLM framework
New open-source LLM optimization framework achieves state-of-the-art results across six domains by treating diverse problems as text artifacts scored by functions.
A new universal optimization framework called optimize_anything treats fundamentally different problem domains—from agent architecture tuning to cloud scheduling to kernel generation—as a single class of text-improvement tasks scored by functions. Researchers Lakshya A Agrawal, Donghyun Lee, Shangyin Tan, Wenjie Ma, Karim Elmaaroufi, and Rohit Sandadi released the work on May 20 as open source through the GEPA project, demonstrating that a single LLM-based system can match or beat specialized tools across six diverse domains.
The framework discovers agent architectures that nearly triple Gemini Flash's ARC-AGI accuracy from 32.5% to 89.5%, finds scheduling algorithms that cut cloud costs by 40%, generates CUDA kernels where 87% match or beat PyTorch performance, and outperforms AlphaEvolve's reported circle packing solution for n=26. It supports single-task search, multi-task search with cross-problem transfer, and generalization to unseen inputs.
What stands out
- 01ARC-AGI accuracy boost: The system improved Gemini Flash's performance from 32.5% to 89.5% on ARC-AGI tasks by discovering improved agent architectures—a 2.75× improvement.
- 02Cloud cost reduction: Scheduling algorithms generated by optimize_anything cut cloud infrastructure costs by 40%, showing the framework can tackle operational problems typically requiring domain-specific solvers.
- 03CUDA kernel generation: 87% of the CUDA kernels produced matched or exceeded PyTorch's performance, indicating the framework can optimize low-level code without specialized compiler knowledge.
- 04Multi-task transfer advantage: Ablation studies revealed that multi-task search outperforms independent optimization when given equivalent per-problem budget, with benefits scaling as more related tasks are added.