Skip to content

About

SwarmEvolve: Latent Agent Swarms for Rapid Evolutionary Search (code coming soon)

Resources

Stars

57 stars

Watchers

2 watching

Forks

Latest commit

 

History

12 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SwarmEvolve

Latent Agent Swarms for Rapid Evolutionary Search

Qile Jiang and George Em Karniadakis
Division of Applied Mathematics, Brown University


Workflow of SwarmEvolve
Step-by-step video with captions (43 s)


SwarmEvolve is a super fast, low-cost, and capable framework for LLM-guided evolutionary search. It solves problems in mathematical discovery and scientific computing by refining and recombining existing programs.

SwarmEvolve uses swarms of open-weight LLM agents, hundreds of them per round, to reason in latent space. Each swarm studies the problem with a different focus and passes their KV caches to a summarizer agent that writes a new program, so the ideas of different agents combine directly in latent space.

Results

Programs evaluated in 60 minutes, mean best score after 15 minutes, and median minutes to reach the 60-minute score of OpenEvolve, with Qwen3 8B
Qwen3 8B, one GPU, 60 minutes per run.

We compare SwarmEvolve with three open-source frameworks:

  • OpenEvolve: A. Sharma, OpenEvolve: an open-source evolutionary coding agent, 2025.
  • ShinkaEvolve: R. T. Lange, Y. Imajuku, and E. Cetin, ShinkaEvolve: Towards Open-Ended and Sample-Efficient Program Evolution, International Conference on Learning Representations, 2026.
  • CodeEvolve: H. Assumpção, D. Ferreira, L. Campos, and F. Murai, CodeEvolve: an open source evolutionary coding agent for algorithmic discovery and optimization, arXiv:2510.14150, 2025.

On 21 problems from mathematical discovery and scientific computing, with Qwen3 8B, 14B, and 32B on one GPU for 60 minutes per run, SwarmEvolve achieves the highest average score with every model.

Best score within 60 minutes for each problem, framework, and model
Best score within 60 minutes, averaged over three runs. Bold: best; underline: second best.

Benchmarks and baselines

  • benchmarks/problems/: the 21 problems, each with task.md, initial_program.py, and evaluator.py.
  • benchmarks/evaluation.py: the shared evaluation used to score every framework.
  • benchmarks/calibration/: expert reference solvers for the scientific-computing problems.
  • baselines/: OpenEvolve, ShinkaEvolve, and CodeEvolve at the commits used in the paper, with the scripts to run them on the benchmarks (setup_baselines.sh, then run_baseline.sh <framework> <problem> <model>).

Code

The SwarmEvolve code will be released soon. For questions or early access, please reach out to the authors.

About

SwarmEvolve: Latent Agent Swarms for Rapid Evolutionary Search (code coming soon)

Resources

Stars

57 stars

Watchers

2 watching

Forks

Releases

Packages

Contributors

Languages