SwarmEvolve: latent agent swarms hit 60-minute baselines in 3-8 minutes on one GPU

burny_tech · x · 2026-11-11

Brown University's Qile Jiang and George Em Karniadakis present SwarmEvolve (arXiv:2611.08483), an evolutionary program-search framework where agents reason in latent space without decoding text.

How it works

450 low-cost agents per round on a frozen Qwen3 8B, each "thinking" 40 steps in hidden-state space
Swarms attack problems from different angles; a novelty filter keeps the 30 most distinct agent trajectories by hidden-state distances
A summarizer reads concatenated KV caches and writes new programs — crossover happens in the cache

Results on 21 math/scientific computing problems

Average best score 0.804 vs 0.693–0.700 for OpenEvolve/ShinkaEvolve/CodeEvolve
Reaches baselines' 60-minute best scores in a median of 3–8 minutes; generates 734 programs/hour vs 81–158
With Qwen3 8B, beats AlphaEvolve's published constructions on three problems

Open question: is the gain from latent reasoning itself or mostly from throughput? Code on GitHub.

Original post →

SwarmEvolve: latent agent swarms hit 60-minute baselines in 3-8 minutes on one GPU

More from coding & agent