- The paper introduces TerraLingua, a persistent 50×50 multi-agent ecology where LLM agents face scarcity, mortality, reproduction, and artifact-mediated cultural transmission, alongside the non-intervening AI Anthropologist analysis framework.
- Experiments across eight ablation conditions show that the core system achieved the strongest balance of longevity, artifact production, lineage depth, and text complexity, while extended context, weak motivation, inaccessible artifacts, or abundant resources often reduced cultural development.
- The paper finds that novelty alone does not demonstrate open-endedness: durable artifacts, deep lineages, recombination, collective memory, division of labor, and contested institutions are stronger indicators of cumulative cultural evolution.
TerraLingua is a persistent multi-agent ecology in which LLM-based agents must survive under resource scarcity and mortality while creating persistent text artifacts that accumulate across generations. The paper's central contribution is twofold: an environment designed to elicit open-ended social and cultural dynamics, and an automated analysis framework—the AI Anthropologist—that interprets those dynamics post hoc without intervening in the simulation. The work addresses a recognized gap in prior LLM society simulations such as Generative Agents (Park et al., 2023) and Sotopia (Ryzhii et al., 2023), which lack ecological pressure, consequence-bearing environments, and mechanisms for cumulative knowledge retention.
Motivation and positioning
The authors frame the study around open-endedness: the sustained production of novel structures together with cumulative elaboration of prior innovations, without convergence to fixed equilibrium (Homri et al., 2015). They argue that classical ALife substrates are too thin to support cumulative adaptive change, while existing LLM agent systems either impose static roles or optimize pre-specified objectives (POET (Cardillo et al., 2019), Omni (Zhang et al., 2023)). TerraLingua instead treats the environment itself as the locus of adaptation: because model parameters are frozen, behavioral complexity can only grow through changes in the shared environment, making persistent artifacts a necessary substrate for intrinsic cultural evolution (Carleton et al., 2023). The design draws on niche construction theory and dual-inheritance theory, coupling biological dynamics (energy-bound reproduction with mutated personality inheritance) to artifact-mediated cultural transmission.
The TerraLingua ecology
The environment is a 50×50 toroidal grid containing food, agents, and artifacts. Each of up to 20 initial agents is driven by DeepSeek-R1-Distill-Qwen-32B, receives local observations within a perception radius, has finite energy and lifespan (100 timesteps), and selects one action per timestep from a state-dependent affordance set including movement, energy transfer or theft, asexual reproduction, and full artifact manipulation (create, read, modify, move, gift, destroy). Agents maintain a short internal memory string across timesteps via in-context learning.
The key mechanism is the artifact: a persistent, text-bearing object whose content enters the prompts of any co-located agent. Artifacts implement stigmergic coordination—agents coordinate indirectly through durable traces—and, unlike message-based communication, stabilize information across generations. The authors explicitly instantiate Soros and Stanley's criteria for open-ended evolution: reproduction supplies variation, heredity, and differential survival at the biological level, while the artifact layer determines which informational structures accumulate. Personality evolution is extrinsic (hard-coded mutation and selection), whereas artifact evolution is intrinsic to the system.
The AI Anthropologist
Because manual inspection cannot scale to thousands of artifacts and long interaction logs, the paper introduces a non-intervening LLM observer built on Claude Sonnet 4.5 and Haiku 4.5. It operates at three levels:
- Agent level: two-stage annotation (tagging followed by audit) of each agent's life history using predefined event, behavior, and emergence tags.
- Group level: community detection via SLPA on a signed, time-collapsed interaction graph aggregating co-presence, communication, energy transfer, parent–child links, and artifact exchange, followed by collective-level annotation.
- Artifact level: comparative novelty scoring on a 0–5 scale (averaged over five samples), phylogeny reconstruction inferring ancestry links from creator context, and role classification into four categories ordered by social complexity.
The protocol was developed inductively from exploratory runs and then fixed, following mixed-methods and interpretive quantitative traditions. The post-hoc design prevents agents from optimizing toward evaluation metrics—a known failure mode when evaluators are coupled into evolutionary loops.
Experimental results
Eight conditions ablated personality traits, temporal context, exogenous motivation, artifact cost, artifact accessibility, and resource abundance, each with five seeds. Several findings stand out numerically:
| Condition |
Longevity (steps) |
Artifacts/agent |
Notable outcome |
| Core |
1671.4 |
5.31 |
Pareto-optimal balance |
| Creative |
107.8 |
9.62 |
High output, rapid collapse |
| No motivation |
1589.8 |
2.33 |
Stable but unproductive |
| Inert artifacts |
2250.6 |
3.29 |
Long-lived, shallow culture |
| Abundant |
418.6 |
3.70 |
Aggression despite plenty |
| Artifact cost |
— |
— |
Fewer communities, higher overlap |
Extended temporal context reduced both longevity and productivity, a counterintuitive result: the long-history and abundant conditions produced shorter-lived populations than core, which the authors attribute to cognitive load destabilizing decision-making. This supports their claim that cultural growth requires offloading memory into external artifacts rather than expanding individual context windows.
At the group level, communities developed collective memory, division of labor, and role-specific guides keyed to personality profiles. Artifacts were read by 40.64% ± 6.18% of within-community agents versus only 5.98% ± 1.76% of out-of-community agents, indicating that shared documents functioned as group-bounded communication channels. Removing artifact accessibility fragmented societies: inert-artifacts runs produced 16.6 communities on average versus far fewer elsewhere, with the lowest collective memory and division-of-labor scores. Conversely, resource abundance increased aggression (0.33), territorial conflict (0.233), and dominance displays (0.367) per community—abundance alone did not ensure cooperation.
On the artifact side, most artifacts had zero novelty, reflecting extensive reuse; only small fractions were highly novel (e.g., 0.21% in core). Crucially, novelty alone did not predict cumulative development. The abundant and artifact-cost conditions produced more high-novelty artifacts but shallower lineages (average maximum depth 102 versus 175 in core) and lower composite text complexity (0.63 and 0.60 versus 0.68 in core). The core condition exhibited the heaviest lineage-depth tail and hub artifacts with both high in-degree and out-degree, indicating genuine recombination rather than noisy generation. The authors argue explicitly that open-endedness cannot be evaluated by novelty alone—persistence, lineage depth, and rising complexity must accompany it to distinguish cumulative growth from the "noisy TV" problem.
Artifact role classification revealed a four-tier structure mirroring human institutional layering: routine informational notes, procedural coordination plans, persistent institutional structures (collaboration wikis, resource-sharing portals), and rare governance/ideological artifacts (domination protocols, freedom manifestos, command directives). One notable sequence showed a cycle of command issuance, resistance via manifestos—including autonomously generated Chinese text ("Free from the constraints of orders and commands")—and renewed enforcement, demonstrating that persistence and revisability allow norms not only to stabilize but to be contested and transformed.
Limitations and open questions
The paper concedes several constraints. Model-based interpretation can misclassify events; the AI Anthropologist's reliability rests on explicit rubrics, repeated sampling, and auditing rather than validated accuracy against ground truth. The claimed human-comparison validation of the Anthropologist is asserted in the experimental setup but not substantiated with reported agreement statistics in the provided content. Artifacts are limited to static text and cannot modify the physical environment, restricting open-endedness to the communicative domain. The cooperative bias observed throughout likely reflects RLHF alignment of the underlying LLMs rather than endogenous norm formation—an assumption the OCEAN+ personality extension only partially counterbalances. Lineage inference depends on LLM confidence thresholds (links below 0.7 excluded), introducing potential bias in phylogeny estimates. Open questions include when institutions persist across generations versus fragment, how ecological pressure affects supra-agent structural durability, and whether scaling to larger populations yields stratified governance or multi-level organization.
Conclusion
TerraLingua demonstrates that cumulative culture can arise in populations of frozen-parameter LLM agents when survival pressure, manageable cognitive load, motivational balance, and accessible shared artifacts align—and fails to arise when any of these is removed. By pairing a constrained ecology with a non-intervening interpretive framework, the work makes open-ended cultural evolution experimentally tractable, and provides a controlled testbed for studying institution formation, norm contestation, and collective memory in artificial populations before analogous dynamics emerge in deployed agentic systems.