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

Kicking off the AIDDA Institute speaker series, Temoor Tanveer (Carnegie Mellon University alumnus, independent researcher) presents LEVI, a framework demonstrating that stronger search architectures can substitute for — or even outperform — larger LLMs in evolutionary search. The talk covers LEVI's conceptual foundations, core design, research findings, and cost-saving mechanisms, and may include a live demonstration on a computational problem. LEVI achieves top benchmark scores at 3.3-6.7x lower computational cost than competing frameworks, matches existing best results on one problem at 35x reduced expense, and outperforms GEPA on prompt optimization benchmarks using less than half the computational budget.