hundreds of low-cost agents per round skip text decoding entirely, passing KV caches between swarm members, with a novelty filter measuring distances between hidden-state vectors to keep ideas distinct. On 21 mathematical-discovery and scientific-computing problems it posted the highest average score and needed a median of only 3–8 minutes to reach the best scores rival frameworks reach in 60 — and with Qwen3 8B it improved on some solutions AlphaEvolve first found with frontier LLMs.