SwarmWorld: Stigmergic technological evolution in societies of language-model agents
Subhadeep Pal, Fiona Y. Wang, Markus J. Buehler
cs.AI, cond-mat.mtrl-sci, cs.CL
2026-08-27
Homogeneous LLMs in SwarmWorld build executable tech without roles. Shared worlds beat isolated search on portfolios and inventions; isolated search still holds the top artifact.
Most multi-agent LLM systems coordinate through chat, assigned roles, or a central workflow. That shows models can talk and specialize. It does not show whether interaction beats the same computational budget spent on independent search. It also rarely asks whether the artifacts still work after their authors are gone.
SwarmWorld turns that into a falsifiable test. Pal, Wang, and Buehler at MIT LAMM drop initially identical language-model agents into a persistent material world with no roles, no recipe catalog, and no global reward. Agents propose architectures and controllers from local observations and private memory. A deterministic simulator decides what can be built and what actually functions.
The world is a lattice of resources, processing stations, and disturbance fields. Agents call gpt-5.6-luna (temperature 0.7) on a fixed staggered schedule, emit schema-checked plans, and have one action applied per tick. Installed artifacts run executable controllers every tick, so later agents can encounter, inherit, and edit them.
Four conditions peel mechanisms apart. Full culture keeps a shared world, messages and records, cross-agent program inheritance, and artifact stigmergy. No communication drops direct speech but keeps shared artifacts and executable inheritance. No explicit culture also drops cross-agent forks, leaving only physical traces. Independent search is N one-agent worlds scored as an endpoint-wise best-of-N envelope, a deliberately strong control.
Discovery and function are scored separately. At checkpoints the state is frozen, all agents are removed, and eight unseen contamination, drought, and storm schedules run on clones. Only physics and installed programs continue. An "invention" must clear a full validation gate: tested materials, a complete design, an installed program, threshold performance, and behavioral novelty.
The 800-tick study crossed N=50/100/200 with four matched seeds. Shared worlds beat the isolated envelope on held-out and portfolio resilience in nearly every cell. At N=200, no-explicit-culture societies added a mean of six validated inventions and a paired discovery-AUC gain of +0.069. Discovery speed was not monotonic: at N=50 full culture still lagged the isolated envelope; at N=100 the shared conditions pulled ahead.
Roles were not prompted. A post-hoc two-cluster model at N=200 put about 27% of full-culture agents in artifact-centered work, versus 20% without explicit culture and 17% without communication. Under full culture, 67%, 76%, and 56% of artifacts at N=50/100/200 recorded multi-agent contributions. Sixteen ranked technologies scored between 0.790 and 0.347, spanning chitin lattices, tidal panels, and mycelial repair membranes rather than one template.
| Setting (N=100, tick 3200) | Portfolio resilience | Validated inventions | Held-out resilience | Best artifact |
| Full culture | 0.2474 | 5.75 | tied with no explicit culture | 0.2380 |
| No explicit culture | 0.2365 | 7.00 | 0.0446 | not separately reported |
| Isolated envelope | 0.1794 | 2.75 | 0.0356 | 0.3488 |
Culture helps portfolios and inheritance. It does not win every endpoint. About 95% of first adoptions began with physical observation, not a chat with the inventor. Randomly removing half the agents left nearly all artifacts tied to a survivor; targeting high-degree nodes hurt far more. The same architecture transferred to a volcanic materials world (AshenRealm) and a sequence-conditioned Protein Realms pilot that installed three protein-matrix materials; the strongest, SELP47K, reached utility 0.729. That pilot is descriptive, not a condition comparison.
For people building multi-agent systems, the useful ruler is not "can they talk." A shared physical substrate already accumulates technology. Explicit culture changes how the population organizes over time; it is not a universal accelerator. Isolated parallel search can still set the single-object record. If the goal is a complementary portfolio that survives after the agents leave, interaction helps. If the goal is one peak score, extra dialogue can add coordination cost.
Each cell has four world seeds, one model, and one prompt. Function is simulator-defined; technology portraits are mechanism visualizations, not manufactured objects. Protein Realms is a single-seed, catalog-conditioned feasibility run. The held-out assay tests programs after agents are removed; the knockout analysis tests graph topology, not physical recovery in a live world. Model weights stay frozen, so there is no online learning.