AIvilization v0: Toward Large-Scale Artificial Social Simulation with a Unified Agent Architecture and Adaptive Agent Profiles

AIvilization v0 is a publicly deployed large-scale artificial society that couples a resource-constrained sandbox economy with a unified LLM-agent architecture, aiming to sustain long-horizon autonomy while remaining executable under rapidly changing environment. To mitigate the tension between goal stability and reactive correctness, we introduce (i) a hierarchical branch-thinking planner that decomposes life goals into parallel objective branches and uses simulation-guided validation plus tiered re-planning to ensure feasibility; (ii) an adaptive agent profile with dual-process memory that separates short-term execution traces from long-term semantic consolidation, enabling persistent yet evolving identity; and (iii) a human-in-the-loop steering interface that injects long-horizon objectives and short commands at appropriate abstraction levels, with effects propagated through memory rather than brittle prompt overrides. The environment integrates physiological survival costs, non-substitutable multi-tier production, an AMM-based price mechanism, and a gated education-occupation system. Using high-frequency transactions from the platforms mature phase, we find stable markets that reproduce key stylized facts (heavy-tailed returns and volatility clustering) and produce structured wealth stratification driven by education and access constraints. Ablations show simplified planners can match performance on narrow tasks, while the full architecture is more robust under multi-objective, long-horizon settings, supporting delayed investment and sustained exploration.

💡 Research Summary

AIvilization v0 presents a publicly deployed, large‑scale artificial society that couples a resource‑constrained sandbox economy with a unified large language model (LLM) agent architecture. The core challenge addressed is the tension between long‑horizon goal stability (teleology) and reactive correctness in a dynamic, rule‑intensive world. To resolve this, the authors introduce three tightly integrated mechanisms.

1. Hierarchical Branch‑Thinking Planner (BTP) – The agent’s overarching life goal is decomposed into multiple parallel objective branches (personal development, production & resource management, trading & market analysis, social engagement). Each branch is further broken down into abstract sub‑tasks. At every planning cycle the agent evaluates internal states (energy, satiety, health, inventory) and external context (prices, eligibility rules) to select the most viable sub‑task, prioritizing those that best serve the high‑level goal and the agent’s personality/value profile. Parallel branches reduce error propagation typical of long sequential plans and enable concurrent reasoning across disparate life domains.

2. Pre‑execution Action Simulator – Before committing actions to the environment, the selected action sequence is forward‑simulated using a predictive world model. The simulator flags resource shortfalls, rule violations, or logical inconsistencies. Upon detection, a two‑tiered repair process is triggered: a lightweight heuristic‑based local fix, and if that fails, a memory‑informed re‑planning step that re‑invokes the hierarchical planner. This design dramatically cuts the need for full re‑planning on every failure, preserving computational budget while maintaining robustness.

3. Dual‑Process Memory – Short‑term memory stores recent execution traces and immediate feedback, supporting rapid correction and adaptation. Long‑term memory consolidates semantic knowledge: values, personality traits, habits, and summarized social records. The long‑term store biases future goal selection and interaction strategies, allowing the agent to maintain a persistent yet evolving identity that can be reshaped by social feedback or human instruction.

The simulated environment incorporates physiological survival costs, multi‑tier non‑substitutable production chains, an automated market maker (AMM) price mechanism, and a gated education‑occupation system. Education imposes minimum entry thresholds; occupations are allocated dynamically based on education level and market conditions, creating realistic scarcity and positional competition.

Empirical evaluation uses high‑frequency transaction data from the platform’s mature phase, constructing 5‑minute OHLC series. Market dynamics exhibit heavy‑tailed return distributions and volatility clustering—stylized facts well documented in empirical finance. Wealth stratification aligns with educational attainment and occupational tiers, reproducing non‑linear wealth‑education relationships observed in labor economics.

Ablation studies compare the full BTP against two simplified variants: “Without‑Branch” (no parallel branches) and “Without‑OD” (no objective decomposition). While simplified planners perform adequately on narrow, single‑objective tasks, they falter in multi‑objective scenarios requiring simultaneous production, earnings, and physiological maintenance. The full architecture excels at coordinated delayed investments (e.g., studying to improve future productivity) and high‑tech production, confirming that hierarchical branching and objective decomposition are essential for robust long‑horizon autonomy.

Human‑in‑the‑loop steering allows users to inject long‑term objectives or short commands. These inputs are stored in the agent’s memory rather than overwriting prompts, ensuring more stable integration of human guidance.

In summary, AIvilization v0 contributes: (i) a unified agent architecture that balances strategic teleology with reactive adaptability under hard constraints; (ii) a richly structured socioeconomic‑education‑labor simulation that yields emergent macro‑phenomena matching real‑world stylized facts; (iii) evidence that hierarchical planning and dual‑process memory are key to robust multi‑objective autonomy; and (iv) a public‑facing platform that can serve as a research‑grade artificial society for studying collective behavior, market dynamics, and institutional effects.