面向大模型辅助模拟版图优化的仿真感知上下文策略改进
文章背景与核心概要
模拟集成电路（IC）版图设计是一个极其耗费人工的过程，严重依赖于迭代式的仿真驱动优化。虽然现有的自动化版图生成器层出不穷，但它们往往无法在没有人工专家介入以及高昂且耗时的后仿真情况下，满足严格的设计规范。
本文介绍了一种仿真感知的大模型多智能体框架，旨在比传统的贝叶斯优化（BO）更高效地优化版图参数。通过在紧凑的结构化版图表示上利用“行动-观察-反思”循环，该框架实现了上下文策略改进（ICPI）。实验结果表明，与内置生成器启发式方法和标准的基于BO的调优相比，该方法表现出更优越的性能，且仅需一小部分后仿真次数。
Simulation-Aware In-Context Policy Improvement for LLM-Aided Analog Layout Refinement
Authors: Bingyang Liu, Ziming Wei, Xiaohan Gao, David Z. Pan
Date: August 13, 2026
Subject: Artificial Intelligence (cs.AI); Robotics (cs.RO)
Publication: Proceedings of the 2026 International Conference on LLM-Aided Design (ICLAD 2026)
Simulation-Aware In-Context Policy Improvement for LLM-Aided Analog Layout Refinement
Authors: Bingyang Liu, Ziming Wei, Xiaohan Gao, David Z. Pan
Date: August 13, 2026
Subject: Artificial Intelligence (cs.AI); Robotics (cs.RO)
Publication: Proceedings of the 2026 International Conference on LLM-Aided Design (ICLAD 2026)
Summary
模拟集成电路（IC）版图设计是一个极其耗费人工的过程，严重依赖于迭代式的仿真驱动优化。虽然现有的自动化版图生成器层出不穷，但它们往往无法在缺乏人工专家介入以及高昂且耗时的后仿真情况下，满足严格的设计规范。
本文介绍了一种仿真感知的大模型多智能体框架，旨在比传统的贝叶斯优化（BO）更高效地优化版图参数。通过在紧凑的结构化版图表示上利用“行动-观察-反思”循环，该框架实现了上下文策略改进（ICPI）。实验结果表明，与内置生成器启发式方法和标准的基于BO的调优相比，该方法表现出更优越的性能，且仅需一小部分后仿真次数。
Summary
Analog Integrated Circuit (IC) layout design is a notoriously labor-intensive process, heavily reliant on iterative, simulation-driven refinement. While automated layout generators exist, they often fail to meet stringent design specifications without manual expert intervention and costly, time-consuming post-layout simulations.
This paper introduces a simulation-aware LLM multi-agent framework designed to optimize layout parameters more efficiently than traditional Bayesian Optimization (BO). By utilizing an "act-observe-reflect" loop on compact, structured layout representations, the framework performs In-Context Policy Improvement (ICPI). Experimental results demonstrate that this approach achieves superior performance compared to both built-in generator heuristics and standard BO-based tuning, requiring only a fraction of the post-layout simulations.
Key Contributions
- 高效优化： 解决了传统贝叶斯优化的样本效率低下问题，传统方法通常需要数百到数千次评估。
- 大模型多智能体框架： 利用大语言模型的推理能力来导航模拟版图复杂的的设计空间。
- 上下文策略改进（ICPI）： 实现了迭代反馈循环，允许模型动态精炼版图优化参数。
- 卓越性能： 在真实世界的模拟电路上进行了验证，该方法在显著降低计算开销的同时，超越了现有的自动化调优启发式方法。
Key Contributions
- Efficient Optimization: Addresses the sample inefficiency of traditional BO, which typically requires hundreds to thousands of evaluations.
- LLM Multi-Agent Framework: Leverages the reasoning capabilities of Large Language Models to navigate the complex design space of analog layouts.
- In-Context Policy Improvement (ICPI): Implements an iterative feedback loop that allows the model to refine layout optimization parameters dynamically.
- Superior Performance: Validated on real-world analog circuits, the method outperforms existing automated tuning heuristics with significantly reduced computational overhead.
Access & Resources
- arXiv ID: 2608.13767
- 全文： 查看 PDF | HTML（实验性）
- DOI： https://doi.org/10.48550/arXiv.2608.13767
Access & Resources
- arXiv ID: 2608.13767
- Full-Text: View PDF | HTML (Experimental)
- DOI: https://doi.org/10.48550/arXiv.2608.13767
Submission History
- [v1] 2026年8月13日 星期四 20:46:17 UTC
Submission History
- [v1] Thu, 13 Aug 2026 20:46:17 UTC