year: 2026/03
paper: https://arxiv.org/abs/2603.16910
website:
code: https://github.com/cognizant-ai-lab/terralingua
connections: cognizant, LLM, multi-agent, simulation, cultural evolution
TerraLingua is a persistent 2D grid world where LLM agents (using DeepSeek-R1-Distill-32B) must survive under resource constraints and limited lifespans. The key twist: agents can create, read, modify, and exchange text-based artifacts that persist in the environment beyond any individual agent’s lifetime.
The main finding is that cumulative culture emerges — agents develop cooperative norms, division of labor, governance structures, and branching artifact lineages — but only when four conditions align: ecological pressure (scarce resources), manageable cognitive load (short context windows actually help), artifact accessibility (agents must be able to read/reuse each other’s artifacts), and balanced motivation (too much creative encouragement destabilizes survival; too little suppresses output).
Counterintuitive results: abundant resources led to more aggression and territorial behavior, not less. Extending agent memory (20 timesteps vs 1) reduced both longevity and artifact production — offloading memory into persistent artifacts worked better than expanding internal context. When artifacts were made invisible (INERT condition), communities fragmented and cultural accumulation stalled, even though populations survived longest.
They also built an “AI Anthropologist” (Claude Sonnet/Haiku) that analyzes the simulation logs post-hoc without intervening — tagging behaviors, reconstructing artifact phylogenies, and scoring novelty. The artifact phylogenies show genuine cumulative lineages: agents build on each other’s artifacts, creating things like energy-sharing protocols, survival guides, and even governance manifestos and counter-manifestos.