LEVI: smarter search replaces costly LLMs, slashing budgets 35x
New framework cuts evolutionary search costs by 3.3-6.7x while matching top results.
LEVI tackles a major inefficiency in LLM-guided evolutionary algorithms: existing frameworks like AlphaEvolve burn through costly frontier-model API calls because they use weak archives, blind model assignment, and full-set evaluation. Instead of throwing larger models at the problem, LEVI redesigns the search architecture itself. It introduces a solution database that preserves diversity from the start, a mutation router that assigns routine edits to small models and hard problems to large ones, and a rank-preserving proxy benchmark that slashes expensive rollouts on redundant examples.
On systems-research benchmarks, LEVI attains the highest score with a budget 3.3–6.7× smaller than published runs of methods like ShinkaEvolve, GEPA, and AdaEvolve. In one case it matched the existing best at 35× lower cost. On prompt optimization, LEVI matches or exceeds GEPA using less than half the rollout budget across four benchmarks. The fully open-source framework shows that smarter architecture, not bigger models, is the path to affordable evolutionary discovery.
- LEVI achieved top scores on systems benchmarks with a budget 3.3–6.7x smaller than ShinkaEvolve, GEPA, and AdaEvolve.
- On one problem, LEVI matched the existing best performance at a 35x lower cost.
- The framework uses a diversity-preserving archive, a smart mutation router (small vs. large LLMs), and a proxy benchmark to reduce costly rollouts.
Why It Matters
LEVI proves stronger search design can slash LLM costs by an order of magnitude, making evolutionary methods practical for more teams.