35B Frontis-MA1 Beats GPT-5.5 via Evolutionary Recursive Self-Improvement

cwolferesearch · x · 2026-08-03

While most recursive self-improvement (RSI) research relies on coding agents for simple hill climbing, the Frontis-MA1 (35B) model introduces evolutionary/genetic algorithms to unlock greater potential.

Trained as a meta-evolution agent for machine learning engineering (MLE), the model features four atomic operators: Draft, Improve, Debug, and Crossover. Instead of generating a single pipeline, it evolves numerous code candidates in a sandbox, learning from real task scores to guide long-horizon search.

On MLE-Bench Lite (constrained to a single RTX 4090), the system boosts the base model's Medal Average from 39.39% to 60.61%. With experience priors and asynchronous search (Evo-Max), it reaches 71.21%, surpassing GPT-5.5 + Codex and approaching GPT-5.6 Sol and the 2.8T Kimi K3.

Original post →

More from coding & agent

coding & agent channel →