LEVI Framework Demonstrates Stronger Search Architectures Can Replace Larger LLMs
Researchers have introduced LEVI, a new open-source evolutionary framework designed to reduce the high computational costs associated with Large Language Model (LLM)-guided evolutionary methods. While existing systems like AlphaEvolve rely on expensive frontier models, LEVI argues that inefficient search architectures, rather than model size, drive up costs. The framework improves three core components: a solution database that maintains diversity, a smart mutation router that optimizes the use of large and small LLMs, and a rank-preserving proxy benchmark. Experimental results show LEVI achieves superior scores on systems-research benchmarks using a budget 3.3 to 6.7 times smaller than competitors such as ShinkaEvolve and GEPA. In one instance, it matched top performance at 35x lower cost. Additionally, LEVI outperformed or matched GEPA in prompt optimization tasks with less than half the rollout budget. This development suggests that architectural improvements can effectively substitute for larger, more costly models in algorithmic discovery and systems research.
Wire timeline
LEVI Framework Demonstrates Stronger Search Architectures Can Replace Larger LLMs
Researchers have introduced LEVI, a new open-source evolutionary framework designed to reduce the high computational costs associated with Large Language Model (LLM)-guided evolutionary methods. While existing systems like AlphaEvolve rely on expensive frontier models, LEVI argues that inefficient search architectures, rather than model size, drive up costs. The framework improves three core components: a solution database that maintains diversity, a smart mutation router that optimizes the use of large and small LLMs, and a rank-preserving proxy benchmark. Experimental results show LEVI achieves superior scores on systems-research benchmarks using a budget 3.3 to 6.7 times smaller than competitors such as ShinkaEvolve and GEPA. In one instance, it matched top performance at 35x lower cost. Additionally, LEVI outperformed or matched GEPA in prompt optimization tasks with less than half the rollout budget. This development suggests that architectural improvements can effectively substitute for larger, more costly models in algorithmic discovery and systems research.
cs.AI updates on arXiv.org