LEVI: Better ADRS Results at a Fraction of the Cost
Summary
This ADRS blog post introduces LEVI, an LLM-based optimization framework designed to cut the cost of AI-driven research for systems (ADRS). It argues that harness-first approaches, using smaller models for most mutations and reserving larger models for paradigm shifts, can achieve stronger results at a fraction of the cost. The post presents benchmark results showing LEVI outperforms baselines and discusses implications for continuous, bespoke optimization in AI research.