← @bgSponge

SwarmWorld: Stigmergic technological evolution in societies of language-model agents

Summary

SwarmWorld shows that homogeneous LLM agents self-organize into technological societies through stigmergy and executable inheritance, producing broader, more resilient portfolios than isolated search.

  1. Cognition is decoupled from physics.

    The simulator enforces energetic constraints and material schemas, while LLM agents propose architectures and controllers within those fixed boundaries.

    SwarmWorld splits cognition from consequence

  2. Shared worlds build resilient portfolios.

    Societies maintain several complementary technologies, whereas isolated agents often excel with only a single strong artifact.

    Societies win by maintaining several complementary technologies

  3. Culture enables executable inheritance.

    Agents do not just discuss ideas; they reuse, edit, and install each other's deterministic, straight-line controllers.

    reused, edited, installed, and propagated one another’s

  4. Roles emerge without explicit prompts.

    Behavioral differentiation between artifact-centered workers and mobile explorers arises naturally as the simulation matures.

    behavioral differentiation arose without role prompts

  5. Stigmergy drives decentralized coordination.

    Agents discover technology through environmental observation and physical changes rather than direct contact with inventors.

    shared physical world acted as an external memory

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →