LEVI: Stronger Search Architectures Can Substitute for Larger LLMs in Evolutionary Search

Researchers introduce LEVI, an open-source evolutionary search framework that achieves superior results on AI research benchmarks while reducing computational costs by 3.3x to 35x compared to existing methods. By optimizing search architecture rather than relying on larger language models, LEVI demonstrates that algorithmic efficiency can significantly reduce the expense of LLM-guided evolutionary discovery.

LEVI addresses a fundamental inefficiency in current LLM-guided evolutionary methods: the assumption that frontier models must power entire search pipelines. Existing frameworks like AlphaEvolve and ShinkaEvolve treat large language models as universal tools, applying expensive frontier-grade inference to tasks ranging from local mutations to comprehensive solution evaluation. This approach drives up computational costs while leaving performance gains unrealized. The research identifies three architectural bottlenecks—poor solution diversity management, undifferentiated model usage, and redundant evaluation—and proposes targeted solutions.

The framework's innovation lies in its harness-first design philosophy, which separates search strategy from model capability. LEVI maintains solution diversity through improved database architecture, routes different mutation types to appropriately-sized models, and implements a rank-preserving proxy system that reduces unnecessary rollouts. These structural improvements enable smaller models to accomplish what previously required frontier-grade inference. The empirical results validate this approach: across systems-research benchmarks, LEVI achieves state-of-the-art performance at a fraction of previous costs, with one benchmark showing 35x cost reduction while matching existing best results.

For the AI research community, this work has immediate implications. As evolutionary search methods proliferate in algorithmic discovery, prompt optimization, and systems research, the ability to achieve frontier-quality results with lower computational budgets expands accessibility and accelerates iteration cycles. The open-source release enables broader adoption and community refinement. The research demonstrates that computational efficiency gains through architectural innovation can rival raw model scaling—a finding relevant to cost-conscious AI development teams.

• →LEVI achieves state-of-the-art results on systems-research benchmarks using 3.3x to 35x less computational budget than existing frontier approaches.
• →Architectural improvements in solution diversity management and mutation routing can substitute for larger language models in evolutionary search tasks.
• →The framework uses rank-preserving proxy benchmarking to reduce redundant rollouts while maintaining search quality.
• →Open-source availability democratizes access to efficient evolutionary search methods previously limited by high computational requirements.
• →Results suggest that algorithmic optimization can be as impactful as raw model scaling for expensive AI research workflows.