LLM‑guided evolutionary search improves programs with AI agents, but each round requires the LLM to generate a long textual chain of thought, making the process slow.\n\nSwarmEvolve deploys hundreds of low‑cost agents per round that reason in latent space and share KV caches, avoiding any decoding to text. Each swarm examines the problem from a different angle, creating diverse search directions.\n\nA novelty filter measures distances between agents' hidden state vectors and retains only the most distinct ideas. Crossover happens naturally in latent space, as a summarizer agent receives the combined KV caches of all swarm members and produces a merged solution.\n\nOn 21 mathematical discovery and scientific computing tasks, SwarmEvolve achieves the highest average score, reaching the best results in just 3 to 8 minutes—far faster than state‑of‑the‑art frameworks that need 60 minutes. With Qwen3 8B, SwarmEvolve further improves solutions that AlphaEvolve initially found on several problems.\n\nThe study shows that agent swarms communicating in latent space offer a fast, low‑cost approach to program search for scientific problems.\n\nReview: SwarmEvolve’s parallel reasoning in latent space dramatically speeds up evolutionary search, highlighting the potential of large models combined with multi‑agent collaboration.