NTH

SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization

AuthorsQile Jiang, George Em Karniadakis

AffiliationsD IVISION OF A PPLIED M ATHEMATICS , B ROWN U NIVERSITY

November 16, 2026 2 min read
Watch on YouTube
The one-line take

SwarmEvolve coordinates large swarms of latent-space AI agents to discover strong scientific programs faster and at lower cost.

Key results

450
Latent agents per round

Agents run across 10 swarms in each evolutionary round.

40
Latent reasoning steps per agent

Each agent reasons in latent space before novelty filtering.

21
Benchmark problems

Mathematical discovery and scientific-computing tasks.

734
Programs evaluated in 60 minutes

SwarmEvolve with Qwen3 8B; OpenEvolve evaluated 158 in the same budget.

0.804
Average best score

SwarmEvolve with Qwen3 8B across the 21 problems.

What the paper found

SwarmEvolve speeds up evolutionary program search by replacing text-based agent reasoning with latent reasoning: each of 450 agents per round takes 40 latent steps, and agents share ideas through transformer KV caches rather than written messages. Across 21 mathematical-discovery and scientific-computing problems, 10 focused swarms explore, refine, and combine candidate programs; a cosine-distance novelty filter preserves diverse reasoning, and each swarm’s summarizer turns the combined caches into a program for evaluation. With Qwen3 8B, the system evaluated 734 programs in 60 minutes, compared with 158 for OpenEvolve, and achieved an average best score of 0.804, ahead of the baselines. It reached competitors’ 60-minute scores in a median of 3 to 8 minutes. Notably, Qwen3 8B also improved on three results associated with DeepMind’s AlphaEvolve, which used frontier Gemini models. The core contribution is a scalable search loop that gets diversity and idea crossover from latent-state communication, avoiding the cost of generating and repeatedly processing long text traces.

Original abstract

LLM-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. We present SwarmEvolve, a framework in which hundreds of low-cost agents per round work in swarms, reason in latent space without decoding to text, and share information by passing KV caches. Each swarm studies the problem from a different angle. A novelty filter keeps the most distinct agent ideas by directly measuring the distances between the agents' hidden state vectors. Crossover of ideas happens naturally in latent space, as a summarizer agent receives the combined KV caches of all swarm members. On 21 problems in mathematical discovery and scientific computing, SwarmEvolve gets the highest average score and needs a median of only 3 to 8 minutes to reach the best scores that other state-of-the-art frameworks reach in 60 minutes. On several problems, SwarmEvolve with Qwen3 8B improves on the solutions that AlphaEvolve first found with frontier LLMs. Our work shows that agent swarms that communicate in latent space offer a fast, low-cost approach to program search for scientific problems.

Read the original paper

More in AI Agents

Browse all 77 papers →