Latent Swarm Intelligence

SwarmEvolve, developed by researchers at Brown University, enables a single GPU to run 450 collaborative AI agents simultaneously, accelerating complex mathematical discovery by orders of magnitude. The framework bypasses traditional text-based communication limitations by using latent embedding space reasoning.

Standard multi-agent systems exchange textual messages that quickly bloat context windows. SwarmEvolve passes transformer key-value caches directly between agents, allowing 40 latent reasoning steps without generating human-readable tokens. Each swarm focuses on specific tasks: mathematical formulation, bug hunting, or literature exploration.

Novelty filtering prevents redundancy by measuring cosine distances between agent hidden state vectors. The system discards the 15 most similar trajectories, retaining the 30 most distinct perspectives. A summarizer agent then synthesizes refined program candidates from the combined caches.

Benchmarked against OpenEvolve, ShinkaEvolve, and CodeEvolve across 21 problems in packing geometry, mathematical analysis, combinatorics, and scientific computing, SwarmEvolve demonstrates significant performance advantages. The latent communication approach eliminates text decoding bottlenecks that typically stall evolutionary search methods.

For operators deploying multi-agent systems, this research suggests practical paths forward. Latent communication reduces computational overhead while maintaining collaborative effectiveness. The 30-perspective filtering mechanism balances diversity with computational efficiency—a critical consideration for production swarms.

The method's success in mathematical domains hints at broader applications. Scientific computing, algorithm optimization, and complex system design could benefit from similar latent-space collaboration frameworks. As multi-agent systems scale, communication efficiency becomes the primary constraint.