Stronger Search Architectures Can Substitute for Larger LLMs (ttanv.github.io)

The recently introduced LEVI framework revolutionizes the approach to code and prompt optimization in AI by demonstrating that more robust search architectures can effectively replace the need for larger language models (LLMs). Rather than incurring high costs by relying on frontier models for evaluations, LEVI optimizes the search process through a tailored architecture that maintains diverse solutions and reduces unnecessary rescoring of examples. This innovative method has shown remarkable results, achieving better performance on various systems-research benchmarks than leading frameworks like GEPA and OpenEvolve, while costing 3.3 to 6.7 times less. LEVI enhances efficiency through three key components: a diversity-preserving solution database, role-aware mutation routing, and a rank-preserving proxy benchmark. By intelligently categorizing mutation calls—sending local refinement tasks to smaller models and reserving larger models for rare structural changes—LEVI minimizes costs while maximizing search effectiveness. Additionally, it constructs a smaller proxy benchmark that maintains candidate rankings without the hefty expense associated with evaluating each potential solution extensively. This combination of innovative architecture and strategic resource allocation positions LEVI as a significant advancement in making AI/ML optimization processes more accessible and cost-effective, potentially accelerating experimentation and innovation within the AI community.