SOCIA-$\nabla$: Textual Gradient Meets Multi-Agent Orchestration for Automated Simulator Generation
This addresses the challenge of brittle prompt pipelines in simulator construction for researchers and practitioners, offering a scalable and reproducible method, though it appears incremental by combining existing multi-agent and optimization concepts.
The paper tackles the problem of automated simulator generation by introducing SOCIA-∇, an agentic framework that optimizes code through a textual gradient descent loop, achieving state-of-the-art accuracy across three CPS tasks such as User Modeling and Personal Mobility.
In this paper, we present SOCIA-$\nabla$, an end-to-end, agentic framework that treats simulator construction asinstance optimization over code within a textual computation graph. Specialized LLM-driven agents are embedded as graph nodes, and a workflow manager executes a loss-driven loop: code synthesis -> execution -> evaluation -> code repair. The optimizer performs Textual-Gradient Descent (TGD), while human-in-the-loop interaction is reserved for task-spec confirmation, minimizing expert effort and keeping the code itself as the trainable object. Across three CPS tasks, i.e., User Modeling, Mask Adoption, and Personal Mobility, SOCIA-$\nabla$ attains state-of-the-art overall accuracy. By unifying multi-agent orchestration with a loss-aligned optimization view, SOCIA-$\nabla$ converts brittle prompt pipelines into reproducible, constraint-aware simulator code generation that scales across domains and simulation granularities. We will release the code soon.