Large Language Models for Agent-Based Social Simulation

33 concepts26 papers9 proposals

Traditional Agent-Based Modeling (ABM) has long been used to study complex systems, from economic markets to disease outbreaks. However, traditional models suffer from a fundamental limitation: their agents are governed by rigid, mathematically simplified rules. An agent might have a fixed probability of passing on a piece of information, or its opinion might be represented as a simple scalar value between -1 and 1 that gets averaged during interactions.

Human society, however, does not operate on scalar averages. Human interaction is mediated by language, context, nuance, and memory. The emerging research direction of LLM-driven agent-based social simulation replaces these rigid mathematical rules with Large Language Models (LLMs) acting as the cognitive engines for individual agents. By allowing agents to reason semantically rather than probabilistically, researchers can simulate human-like social dynamics with unprecedented fidelity.

The Shift from Scalar Rules to Semantic Reasoning

To understand why LLM-driven simulation works, consider how information spreads. In a traditional epidemiological model of information propagation, a message is treated like a static virus. But as Yuxuan Hu 0002 et al. demonstrate with the LAID framework in An LLM-enhanced Agent-based Simulation Tool for Information Propagation, real information evolves as it diffuses. Humans interpret, filter, and actively modify messages based on their unique personalities and interests. LLM agents achieve this because they perform semantic reasoning on received text, altering the content before passing it along.

This shift from numeric variables to natural language reasoning fundamentally changes how collective behavior emerges. For example, when modeling political polarization, traditional models struggle to explain why opinions harden. In Decoding Echo Chambers: LLM-Powered Simulations Revealing Polarization in Social Networks, Chenxi Wang et al. replace scalar opinion models with language-based reasoning. Because the agents process information through a complex persona, they don't just average out opposing views; instead, they exhibit realistic human behaviors like opinion entrenchment and linguistic justification.

The Architecture of a Social Agent

For an LLM to act effectively as a social agent, it cannot simply be a stateless text predictor. It requires an architectural framework that mimics human cognitive constraints. As surveyed by Xinyi Li et al. in A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges, modern LLM-based multi-agent systems standardize agent design into several core modules:

  1. Profile and Persona: Agents are assigned explicit biographies, backgrounds, or biases. Yun-Shiuan Chuang et al. leverage this in The Wisdom of Partisan Crowds: Comparing Collective Intelligence in Humans and LLM-based Agents by instantiating partisan personas. They found that these role-playing agents successfully replicate human-like biases while retaining the ability to improve collective estimates through deliberation.
  2. Memory and Reflection: To maintain consistency over long timelines, agents need to store past interactions. Roberto Gallotta et al. note in Large Language Models and Games: A Survey and Roadmap that persistent memory and reflection architectures are what allow agents to plan collective events and maintain narrative consistency across multiple interactions.
  3. Perception and Action: Agents must perceive unstructured environmental cues and translate their internal reasoning into actions. In the RecAgent framework proposed by Lei Wang et al. in User Behavior Simulation with Large Language Model-based Agents, this loop enables agents to browse web platforms, generate contextually relevant reviews, and accurately mirror human user conformity.

From Micro-Reasoning to Macro-Phenomena

The true power of LLM-driven simulation is that micro-level semantic reasoning scales up to reproduce macro-level social phenomena. This makes it an invaluable tool for computational social science, as highlighted in the comprehensive survey by Taicheng Guo et al. in Large Language Model Based Multi-agents: A Survey of Progress and Challenges.

Consider a crisis scenario like a natural disaster. Traditional simulations rely on strict "if-then" evacuation routes. However, in RESPOND: Realistic Environment Simulation of Population and Natural Disasters with LLM-Driven Agents, Roman Sultimov et al. demonstrate that when agents are given LLM "brains," they can process unstructured text from social media updates alongside official alerts. If an official alert contradicts what an agent perceives on social media, the agent uses context-aware reasoning to prioritize personal safety goals over official instructions—perfectly capturing the messy, unpredictable nature of real human compliance during crises.

Prototyping and Sandboxing Human Environments

Because these agents can mimic human cognitive processes—even in chaotic environments like relationship pairing, as explored by Sihan Chen et al. in CogniPair - Dynamic LLM Matching Algorithm in Chaotic Environments Mimicking Human Cognitive Processes for Relationship Pairing—they provide a safe, ethical sandbox for testing social hypotheses that are impossible to conduct in the real world.

Researchers are deploying these cognitive sandboxes across diverse domains:

Ultimately, as Tianshi Li 0001 et al. argue in Shaping the Emerging Norms of Using Large Language Models in Social Computing Research, this paradigm of "synthetic social simulation" allows the scientific community to prototype complex social systems and replicate foundational social science studies rapidly. While challenges remain—such as the performance degradation and honesty issues identified by Yixuan Zhang et al. in the MetaAgents: Large Language Model Based Agents for Decision-Making on Teaming framework—grounding agent behavior in natural language reasoning marks a permanent shift away from rigid, rule-based digital worlds toward fluid, human-like digital societies.

Key Foundational Work

AI for social science and social science of AI: A Survey (2024) — contributed 2 concept elements to this direction.
SurgBox: Agent-Driven Operating Room Sandbox with Surgery Copilot (2024) — contributed 2 concept elements to this direction.
A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges (2024) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_llm_driven_agent_reasoning_9524b11d,
  title = {LLM-driven agent reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/llm-driven-agent-reasoning/rdr_9524b11d49404f048b24fa6ac17d7c95},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_llm_driven_agent_reasoning_9524b11d_v1,
  title = {LLM-driven agent reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/llm-driven-agent-reasoning/rdr_9524b11d49404f048b24fa6ac17d7c95/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake