OPEN LINK ↗
// 75d agoBENCHMARK RESULT
LEVI beats GEPA, OpenEvolve on budget
LEVI is a new LLM-guided evolutionary optimization framework that emphasizes search-harness design over frontier-model dependence, using cheaper models for most mutations and stronger models only for paradigm shifts. Based on its published ADRS and circle-packing results, it reports better or tied performance against GEPA/OpenEvolve baselines at significantly lower cost and has released code on GitHub.
// ANALYSIS
The interesting claim here is not just higher scores, but that architecture and diversity strategy can outperform brute-force model spend in evolutionary coding loops.
- –LEVI’s stratified model allocation pushes most mutation traffic to cheaper Qwen-class models, reserving expensive calls for rare high-creativity steps.
- –Its fingerprint-based CVT-MAP-Elites combines structural and performance diversity, which helps keep search breadth without collapsing too early.
- –Controlled comparisons on the same model and evaluation budget suggest the gain is in search mechanics, not simply model quality.
- –If these results replicate broadly, LEVI could lower the cost floor for ADRS-style optimization and make this workflow practical for smaller teams.
// TAGS
levillmbenchmarkresearchopen-source
DISCOVERED
75d ago
2026-03-14
PUBLISHED
76d ago
2026-03-12
RELEVANCE
9/ 10
AUTHOR
Longjumping-Music638