LEVI: cheaper, faster LLM optimization framework
LEVI, a new LLM‑based optimization framework, claims to outperform prior AI research systems while cutting cost 3–7x by using cheaper models for mutations and stronger frontier models for paradigm shifts — and it ships code and a writeup. That model-of-models approach could shift how small teams explore architecture search or automated research loops. (x.com)
ttanv/levi is published as a public GitHub repository (MIT license) with ~285 commits, roughly 10–11 stars and 2 forks on the project page. (t.co) The code exposes a single high-level API entry levi.evolve_code and accepts model= or paradigm_model=/mutation_model= arguments, plus budget_dollars, budget_evals, or budget_seconds for run-level cost control; example entrypoints live in examples/circle_packing and examples/ADRS. (github.com) The write‑up describes a “stratified model allocation” that routes cheap local models to frequent refinement (mutations) and reserves expensive frontier models for occasional paradigm shifts, and it adapts MAP‑Elites style quality‑diversity (CVT‑MAP‑Elites) together with GEPA‑inspired Pareto ideas for diversity maintenance. (ttanv.github.io) Concrete ADRS claims appear in the README: LEVI reports the highest ADRS benchmark scores among compared frameworks in controlled equal‑budget experiments (same model, 750 evals) and faster convergence than baselines. (github.com) A specific example listed: on the n=26 circle‑packing ADRS task LEVI logged a packing density of 2.6359+ on a $15 budget using local mutation models Qwen3‑30B‑A3B and Xiaomi MiMo‑v2‑Flash with Gemini 3 Flash used for periodic paradigm shifts, and the README notes the majority of accepted mutations came from the local Qwen model. (github.com) The project frames cost wins numerically, stating a representative run cost of $4.50 versus $15–$30 claimed for other frameworks, and includes a quickstart that shows a minimal example using openai/gpt‑4o‑mini with budget_dollars=2.0 to demonstrate low‑dollar experimentation paths. (github.com)
