SwarmEvolve uses latent agent swarms for faster evolutionary search
Original title: SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization
SwarmEvolve reaches top scores on 21 problems in 3-8 minutes using latent-space agent swarms.
LLM-guided evolutionary search is slow because each LLM call generates long text chains and must repeat generations to preserve diversity. SwarmEvolve deploys hundreds of low-cost agents that reason in latent space without decoding to text and share KV caches. Each swarm studies the problem from a different angle with a novelty filter measuring hidden-state distances. Crossover occurs naturally via a summarizer agent combining KV caches. On mathematical discovery and scientific computing tasks, SwarmEvolve achieves the highest average score and is faster than state-of-the-art frameworks.
- Median time 3-8 minutes to reach best scores
- Hundreds of low-cost agents per round
- Novelty filter keeps most distinct agent ideas
- Crossover in latent space via KV caches
- Improves on AlphaEvolve solutions with Qwen3 8B
From the original article
1 min read at arxiv.orgLLM-guided evolutionary search uses AI agents to improve programs for mathematical and computational problems by iteratively refining solutions and combining high-scoring methods. However, the process is often slow because the LLM generates a long textual chain of thought for each round and must repeat this generation many times to maintain population diversity.
By Qile Jiang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.