Behavior Generation for Heterogeneous Agents in Urban Simulation Deduction: A Multi-Stage Approach Based on Large Language Models

Traditionally, constructing a professional simulation deduction script is a long-cycle process that requires collaboration between domain experts and simulation experts. The increasing demands for simulation deduction across large-scale scenarios and emergency situations highlight the need for rapid simulation scenario construction methods. Recent advances in large language models (LLMs) offer powerful new tools to address these challenges. In this work, we propose LLM-HABG, a highly portable framework that can automatically generate behavior plans for heterogeneous agents in urban simulation environments based on natural language instructions, requiring no additional post-training. It comprises four stages: task extraction, task decomposition and allocation, urban information retrieval, and behavior generation. To overcome the limitations of LLMs, such as deficient instruction-following and hallucinations, we incorporate a suite of enhancement techniques including dynamic few-shot learning, entity alignment, and constrained decoding. Experimental results demonstrate that LLM-HABG achieves robust performance even with 14B-parameter models, attaining a success rate exceeding 75 %-nearly doubling baseline performance. This conversational behavior configuration paradigm empowers domain experts to interact directly with simulation systems via natural language, significantly reducing the cost and barriers associated with simulation scenario construction.