arXiv cs.AI· Yuncheng Hua, Sion Weatherhead, Mehdi Jafari, Hao Xue, Flora D. Salim·· 11 小时前AI 评分32
SOCIA-∇：文本梯度结合多智能体编排实现自动化模拟器生成
SOCIA-$\nabla$: Textual Gradient Meets Multi-Agent Orchestration for Automated Simulator Generation
AI 导读
SOCIA-∇ 是一个诊断驱动的多智能体框架，用于自动化构建模拟器代码，让 LLM 智能体在文本计算图中迭代完成合成、执行、评估、诊断与修复。该框架在 LLMob 个人移动性基准和半合成地理社交口罩采纳基准上，于分布内模拟、分布外 what-if 生成和地理社交扩散三类任务中取得最强整体模拟器保真度。代码与数据已公开，论文被 ACM SIGSPATIAL 2026 接收。
正文 · 原文
Abstract:In this paper, we present SOCIA-$\nabla$, a diagnosis-driven multi-agent framework for automated simulator-code construction. SOCIA-$\nabla$ iteratively refines executable simulator code within a textual computation graph, where LLM-driven agents synthesize, execute, evaluate, diagnose, and repair simulators. A grounded simulator blueprint attributes behavioral residuals to candidate mechanisms and their spatial or interaction contexts, which are then converted into localized textual repair signals with history-aware memory and constraint projection. We evaluate SOCIA-$\nabla$ on the real-world LLMob personal-mobility benchmark and a controlled semi-synthetic geo-social mask-adoption benchmark. Across in-distribution simulation, out-of-distribution what-if generation, and geo-social diffusion, SOCIA-$\nabla$ achieves the strongest overall simulator fidelity. SOCIA-$\nabla$'s code and data are publicly available at this https URL.
| Comments: | 14 pages, 2 figure, 3 tables. This paper has been accepted to the ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems 2026 (ACM SIGSPATIAL 2026), the thirty-fourth edition |
| Subjects: | Artificial Intelligence (cs.AI) |
| ACM classes: | I.2.7 |
| Cite as: | arXiv:2505.12006 [cs.AI] |
| (or arXiv:2505.12006v5 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2505.12006 arXiv-issued DOI via DataCite |
Submission history
From: Yuncheng Hua [view email]
[v1]
Sat, 17 May 2025 13:47:31 UTC (3,517 KB)
[v2]
Fri, 23 May 2025 15:30:42 UTC (2,940 KB)
[v3]
Tue, 21 Oct 2025 13:39:58 UTC (1,626 KB)
[v4]
Tue, 11 Nov 2025 06:15:00 UTC (2,673 KB)
[v5]
Wed, 4 Nov 2026 23:20:06 UTC (3,079 KB)
来源：arXiv cs.AI · arxiv.org