SwarmWorld: Stigmergic technological evolution in societies of language-model agents
Homogeneous LLM agents built durable simulated technologies through shared artifacts, not assigned roles or central control.
SwarmWorld has agents explore, process resources, test materials, build artifacts, and write controllers that are later judged by a deterministic simulator under unseen disturbances. The paper says shared societies produced broader and more resilient portfolios than a best-of-N isolated-search baseline, though isolated search still competed on the single strongest artifact. Agents differentiated into exploration, construction, maintenance, and coordination as the world matured. Most reuse began through physical observation rather than communication, with explicit cultural mechanisms helping organization but not guaranteeing better inventions. ArXiv · AI/CL/LG's note
SwarmWorld has agents explore, process resources, test materials, build artifacts, and write controllers that are later judged by a deterministic simulator under unseen disturbances. The paper says shared societies produced broader and more resilient portfolios than a best-of-N isolated-search baseline, though isolated search still competed on the single strongest artifact. Agents differentiated into exploration, construction, maintenance, and coordination as the world matured. Most reuse began through physical observation rather than communication, with explicit cultural mechanisms helping organization but not guaranteeing better inventions. ArXiv · AI/CL/LG's note
score 5