Google Research's SwarmEvolve Runs 450 Latent-Reasoning Agents, Beats AlphaEvolve on 3 Problems With Qwen3-8B

SwarmEvolve replaces text reasoning with latent-space messaging via KV-cache transfer across 10 swarms of 45 agents, evaluating 734 programs per hour versus 158-402 for baselines. Running Qwen3-8B, it reaches the baseline 60-minute score in 3-8 minutes and surpasses AlphaEvolve's Gemini-driven solutions on three problems across its 21-task suite covering math discovery and scientific computing.