SwarmEvolve: Hundreds of Latent-Space Agents Hit OpenEvolve's 1-Hour Scores in 5 Minutes

QileHQ · reddit · 2026-11-09

Inspired by DeepMind's AlphaEvolve (which broke the 1969 record for 4x4 complex matrix multiplication via LLM-driven evolutionary search), Brown University's Qile Jiang and George Em Karniadakis present SwarmEvolve, a faster, cheaper, scalable framework for mathematical discovery and optimization.

How it works

Runs swarms of hundreds of latent-reasoning agents on Qwen3 8B/14B/32B;
Agents never write out thoughts: inspired by LatentMAS, they reason in latent space and pass KV caches to each other, mixing ideas inside the model for massive throughput gains;
Small agents are cheap, so many run in parallel with different focuses, giving more diverse candidates per round;
A novelty filter measures diversity via distances between agents' hidden state vectors, keeps diverse agents, and with 50% probability prompts new agents with the best program found so far.

Results: reaches in 5 minutes what OpenEvolve, ShinkaEvolve, and CodeEvolve need >1 hour to achieve; with just the 8B model it matched the best known circle packing; extended to Qwen3.5 and Muse models. Authors invite collaborations on real scientific discovery problems.

Original post →

SwarmEvolve: Hundreds of Latent-Space Agents Hit OpenEvolve's 1-Hour Scores in 5 Minutes

More from coding & agent