Researchers released LEVI, a new framework for AI‑Driven Research for Systems that slashes the cost of algorithmic discovery. By delegating most code mutations to a modest QWEN 30B model and reserving larger LLMs for occasional paradigm shifts, LEVI keeps the search archive diverse across structure and behavior. Benchmarks show it delivers comparable or better results while spending far less than traditional.

Current ADRS pipelines rely on expensive, closed‑source LLMs, limiting accessibility and preventing continuous optimization. The authors argue that ADRS should evolve into a CI/CD‑like process, re‑optimizing algorithms whenever workloads, hardware, or service‑level objectives change. Lowering the price tag to roughly 3–7× cheaper than existing baselines makes bespoke, per‑deployment tuning feasible for enterprises across global data‑center environments today and in real‑time.

By focusing on the search harness rather than always using the biggest model, LEVI maintains both structural and behavioral diversity, preventing the archive from collapsing into a single solution family. This approach lets researchers generate volumes of incremental variants cheaply while still injecting occasional, high‑impact ideas. The result is a more practical, scalable path for deploying custom‑tuned algorithms at production scale.