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

TL;DR

AIvilization v0 tackles the challenge of sustaining teleologically coherent yet adaptively correct behavior for large-scale LLM-driven agents in a resource-constrained artificial society. It introduces a unified cognitive core featuring a Branch‑Thinking Planner, pre-execution Action Simulator, adaptive Dual-Process Memory, and human-in-the-loop steering to manage long-horizon objectives under dynamic constraints. The platform couples physiological survival, a Leontief-like production network with non-substitutable inputs, and an AMM-based price mechanism to generate emergent macro phenomena such as inflation signals and wealth stratification, which align with canonical stylized facts. Ablation studies reveal that hierarchical branching and objective decomposition improve robustness in complex multi-objective tasks, while lighter planning suffices for simple tasks, underscoring the system’s flexibility and potential as a research-grade testbed for emergent social dynamics and hybrid governance.

Abstract

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.

Figures (14)

• Figure 1: Internal organization of the agent cognitive core and its interaction with the environment. The cognitive system is decomposed into four functional components: (A) a dual-process memory system supporting consolidation, and working context; (B) a hierarchical planner that transforms abstract goals into primitive actions through horizon, branch, and action layers; (C) an action simulator that filters and validates proposed actions; and (D) a simulation environment that executes actions and generates experience. Information flows form a closed feedback loop linking memory, planning, execution, and learning.
• Figure 2: Agent cognitive architecture with hierarchical planning, simulation-based filtering, and dual-process memory. Overview of the agent-level cognitive system governing planning, execution, and adaptation. Long-term memory and short-term memory jointly provide guidance and working context to a hierarchical planner, which decomposes abstract goals into branches and executable actions. Proposed actions are evaluated by an action simulator to detect infeasible or conflicting behaviors, enabling lightweight repair or full re-planning before execution in the environment. Environmental feedback and execution outcomes are fed back to update short-term state and consolidate long-term experience.
• Figure 3: Economic environment and production–consumption structure of the simulated world. Schematic of the closed economic system in which agents operate. Production is governed by a resource-constrained production function combining time, energy, satiety, residential level, and materials. Goods flow through a vertically structured supply chain spanning primary resources, consumer goods, intermediate inputs, and high-technology sectors. Agents simultaneously act as producers and consumers, participating in labor, production, trading, and consumption cycles that jointly determine market dynamics and resource allocation.
• Figure 4: High-frequency price and volume series from the simulated market, covering the period 2025-09-09 to 2025-09-15 in real-world time, where the simulation operates at a time compression ratio of 7:1 relative to real-world time. Intraday candlestick prices is at the top and corresponding traded volume is at the bottom.
• Figure 5: Time series of normalized price multipliers for three stages of the silicon value chain: raw ore, intermediate silicon, and finished transistors.