ABSClaw: From Invented Worlds to Reality-Grounded Social Simulation

Shaobin He, Yan Gao, Quan Zheng, Zhenyang Lin, Zhen Liu, Jiyan He, Shuxin ZhengarXiv 2026 · arXiv

TL;DR

ABSClaw is a zero-code framework that transforms natural-language requests into scientifically grounded social simulations. By grounding agents in real-world data and compiling simulation logic into a fixed vocabulary of 26 atomic operations, it creates more reliable, reproducible, and massive-scale artificial societies than previous free-form code generation methods.

What Problem Does This Paper Address?

Large Language Model (LLM) agents have made it easy to build "artificial societies," but these simulations often suffer from a lack of scientific grounding. The authors identify three major gaps:

  1. Invented Worlds: Most simulations use "toy" personas, generic environments, and synthetic social networks that don't reflect real populations.
  2. Brittle Implementation: Existing systems often ask an LLM to generate the entire simulator as free-form code. This makes results highly sensitive to the specific LLM used (the "backbone model") and difficult to reproduce.
  3. Hard-wired Capabilities: Simulators are often built for specific tasks (like public health or urban mobility) and cannot easily be extended with new data sources or analysis tools without a total redesign.

Methods

ABSClaw addresses these issues through a four-step compilation pipeline:

  1. Requirement Capture: Resolves a user's natural language request into a structured specification.
  2. Real-World Grounding: Instead of inventing data, it "materializes" the world using real sources. It pulls personas from census data and the World Values Survey (WVS), environments from OpenStreetMap and weather APIs, and social structures from empirical network datasets.
  3. Atomic Operation Compilation: This is the core innovation. Instead of writing unconstrained code, ABSClaw compiles the simulation logic into a program using a fixed vocabulary of 26 atomic operations (e.g., create_agent, send_message, llm_decide). This "constrained" approach ensures the simulation structure is consistent regardless of which LLM is writing the code.
  4. Simulation Execution: The program runs on a shared runtime, producing execution traces, metrics, and reports.
Figure 1: ABSClaw workflow.

The system is also designed to be extensible through "Skills"—modular plugins that can add new data accessors or analysis capabilities without touching the core simulator.

Datasets

The authors introduce ABSBench, the first systematic benchmark for evaluating executable social-simulation construction.

  • ABSBench: Contains 193 real-world tasks across ten domains, including urban mobility, health behavior, and crisis management.
  • Grounding Sources: The framework utilizes the World Values Survey (WVS), SNAP (Stanford Network Analysis Project) for social topologies, OpenStreetMap (OSM) for geospatial data, and Open-Meteo for historical weather.
Composition of the 193 ABSBench tasks across domains, granularity, and difficulty.

Key Insights and Ideas

  • Constraint as a Feature: By forcing the LLM to use a specific intermediate representation (the 26 atomic operations), the authors significantly reduced "backbone sensitivity." This means that even smaller, less powerful LLMs can generate high-quality simulations that are comparable to those generated by top-tier models.
  • Three-Axis Alignment: Grounding who acts (personas), where they act (environment), and with whom they interact (networks) independently and cumulatively improves how well the simulation matches real-world outcomes.
  • Zero-Code Scalability: By integrating a specialized "runtime skill" (Adaptive Prototype Simulation), ABSClaw was able to scale from small groups to 25 million agents while maintaining a zero-code interface for the user.

Key Results

  • Superior Performance: On the ABSBench benchmark, ABSClaw achieved a score of 168.59/200, outperforming other state-of-the-art systems like SOCIA-EVO and OneSim.
  • Reduced Model Sensitivity: ABSClaw reduced performance variation across different backbone LLMs by 61.6% compared to direct code generation.
  • Real-World Fidelity:
    • In a case study of the 2026 Scottish Parliament election, grounding personas in census data reduced vote share prediction error by 46.4%.
  • In a road traffic simulation during a major storm in Finland, environment grounding improved flow-ratio accuracy significantly.
  • Massive Scale: The system successfully executed 13 population-scale simulations, collectively involving 25 million agents.
Comparison of ABSClaw performance against other systems across different task strata.

Limitations and Future Work

While ABSClaw significantly improves the construction and grounding of simulations, the authors note that "plausible role-playing does not guarantee valid collective behavior." The macro-level outcomes of a simulation still depend heavily on environmental dynamics and how agents are scheduled.

Future work includes expanding the library of "Skills" to cover more niche social science theories and further investigating how to align individual LLM decision-making with complex, long-term social trends.

Summary updated

Abstract

LLM-based artificial societies exhibit rich behavior but often lack scientific grounding: personas and environments are assumed, interaction networks are synthetic, and simulators are generated as free-form code, making results difficult to tie to real populations and sensitive to the backbone model. We argue that simulated worlds should be sampled from observed data, while simulation logic should be implemented by composing operations from a shared, task-independent vocabulary. In this paper, we introduce ABSClaw, a zero-code agent that compiles natural-language requests into executable, reality-grounded social simulations. ABSClaw grounds who acts, where they act, and with whom they interact in survey and census data, geospatial sources, and real networks. It expresses state transitions through 26 atomic operations. To support an open and extensible ecosystem, ABSClaw further allows capabilities to be added as lifecycle-scoped skills without modifying its core. Furthermore, we propose ABSBench, a benchmark of 193 real-world social-simulation tasks. On ABSBench, ABSClaw achieves the best overall performance among the evaluated social simulators. Three real-world case studies show that grounding personas, environments, and interaction networks each improves agreement with observed outcomes. With one runtime skill, ABSClaw executes 13 population-scale simulations, collectively involving 25 million agents. Together, these results demonstrate ABSClaw's real-world fidelity, extensibility, and scalability.
BibTeX
@misc{he2026absclaw,
  author = {Shaobin He and Yan Gao and Quan Zheng and Zhenyang Lin and Zhen Liu and Jiyan He and Shuxin Zheng},
  title = {{ABSClaw: From Invented Worlds to Reality-Grounded Social Simulation}},
  year = {2026},
  eprint = {2611.24517},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.24517},
}
Figure