SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization

SwarmEvolve. Latent Agent Swarms for Evolutionary Discovery and Optimization. By Qile Jiang and one other, from Brown University. Published on the fifth of November.

In LLM-guided evolutionary search, AI agents improve programs for mathematical and computational problems. They refine solutions round after round, and they combine methods that score well, much as an evolutionary algorithm breeds its fittest members. The process tends to be slow, for two compounding reasons. In each round the LLM writes out a long textual chain of thought. And to keep the population diverse, it has to repeat that generation many times over. Long outputs, multiplied by many repetitions, multiplied by many rounds: the text itself becomes the bottleneck.

SwarmEvolve takes the text out of the loop. Each round runs hundreds of low-cost agents organised into swarms. They reason in latent space, in the model's internal representations, and never decode their thinking into words. They share information with one another by passing KV caches, rather than by writing messages. Each swarm approaches the problem from a different angle, which is where the population's diversity now comes from, instead of from regenerating long text again and again.

Two evolutionary operations carry over into this latent setting. Selection becomes a novelty filter. It measures distances between the agents' hidden state vectors directly and keeps the ideas that are most distinct, with no need to read them as text first. Crossover, the combining of ideas, needs no separate step either. A summarizer agent receives the combined KV caches of every swarm member, and the ideas mix there, in latent space, as a natural consequence of being read together. Variation, selection and recombination all happen without a word being written out.

Testing covers twenty-one problems in mathematical discovery and scientific computing. Across them, SwarmEvolve gets the highest average score. Then there is time. Other state-of-the-art frameworks reach their best scores in sixty minutes, and SwarmEvolve matches those scores in a median of only three to eight minutes.

A further result concerns earlier discoveries. On several problems, SwarmEvolve running Qwen3 8B improves on solutions that AlphaEvolve first found using frontier LLMs.

Evolutionary program search had been paying for its diversity in text: long chains of thought, generated again and again. Move the agents' reasoning and their communication into latent space, and that bill shrinks. The authors' conclusion is that swarms talking in latent space give a fast, low-cost approach to program search on scientific problems.