Integrating LLM in Agent-Based Social Simulation: Opportunities and Challenges скачать в хорошем качестве

Integrating LLM in Agent-Based Social Simulation: Opportunities and Challenges

This paper explores the *integration of Large Language Models (LLMs) into social simulation**, evaluating both their promising potential and significant limitations from a computational social science perspective. LLMs, trained on vast amounts of human text, can generate human-like language and behavior, demonstrating impressive abilities in tasks such as the Turing Test and Theory of Mind assessments. However, this capacity stems from **statistical pattern recognition rather than genuine understanding or consciousness**, leading to critical issues like **cognitive biases, factual inconsistencies (hallucinations), lack of behavioral diversity (converging to an "average persona"), and a "black-box" nature* that makes their reasoning difficult to interpret. While LLMs are valuable for *interactive simulations, training, and exploratory modeling* due to their flexibility and ability to create believable experiences, their use in *scientific, explanatory, or predictive social simulations is more challenging* due to these inherent limitations, high computational costs, and difficulties in validation. Therefore, the paper advocates for *hybrid approaches that combine the expressive flexibility of LLMs with the analytical rigor and transparency of traditional rule-based agent-based modeling (ABM) platforms* like GAMA and NetLogo, suggesting a future where LLMs are powerful components within a broader, more sophisticated modeling ecosystem. https://arxiv.org/pdf/2507.19364