From the source
The paper introduces a system that coordinates multiple LLMs by dynamically assigning roles (Thinker, Worker, Verifier) using a lightweight coordinator with fewer than 20K learnable parameters, trained via evolutionary algorithms.
The system achieves state-of-the-art results on LiveCodeBench (86.2% pass@1) and generalizes zero-shot to unseen tasks, outperforming individual models including GPT-5, Gemini 2.5-Pro, and Claude-4-Sonnet.