SwarmEvolve uses latent-agent swarms on Qwen3 to match OpenEvolve results in 5 minutes

·2026.11.10 00:30

Key point

The framework runs hundreds of latent-reasoning agents on Qwen3 8B, 14B, and 32B models to reach scores that took other frameworks over an hour.

SwarmEvolve is a new framework for math discovery and optimization that uses swarms of latent-reasoning agents to improve throughput and cost efficiency.
Agents reason in latent space and pass KV caches to each other, allowing ideas to mix within the model without writing out explicit thoughts.
The system reaches scores in approximately 5 minutes that took OpenEvolve, ShinkaEvolve, and CodeEvolve more than 1 hour to achieve.
Using only the Qwen3 8B model, the framework matched the best known circle packing results.
The framework has been extended to work with Qwen3.5 and Muse models, showing even better performance.
Code is available on GitHub, and the paper is hosted on alphaxiv.org.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.