SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization
SwarmEvolve reports reaching rivals’ hour-long search results in a median of 3–8 minutes.
Across 21 mathematical discovery and scientific computing problems, the framework achieved the highest average score in the authors’ comparison. Hundreds of agents per round reason without generating text, sharing internal model states through KV caches. A novelty filter retains distinct ideas, while a summarizer combines swarm members’ caches to produce new candidates. On several problems, SwarmEvolve using Qwen3 8B improved on solutions first found by AlphaEvolve with frontier models.
HF Daily Papers' note
Across 21 mathematical discovery and scientific computing problems, the framework achieved the highest average score in the authors’ comparison. Hundreds of agents per round reason without generating text, sharing internal model states through KV caches. A novelty filter retains distinct ideas, while a summarizer combines swarm members’ caches to produce new candidates. On several problems, SwarmEvolve using Qwen3 8B improved on solutions first found by AlphaEvolve with frontier models.
HF Daily Papers' note
score 5