Ilmu Komputer & AI editorial
Open AccessOA2026
AgenticSizing: A Large Language Model-based Multi-Agent Framework for Analog Circuit Sizing
A multi-agent LLM framework that decomposes analog circuits into functional blocks, extracts reusable design knowledge, and coordinates specialized agents to meet global performance specifications.
Yijia Hao; Pratibha Verma; Dongxu Guo; Cristian Sestito; Michael O'Boyle; Christos-Savvas Bouganis; Themis Prodromakis· 2026· DOI 10.48550/arXiv.2609.25873
The core problem
Analog circuit sizing remains a challenging and time-consuming task due to the large design space, strong performance trade-offs, and increasing circuit complexity in scaled technologies. Recent large language model (LLM)-based methods show promise in improving sample efficiency and interpretability, but existing approaches often lack explicit circuit-topology understanding and are mainly evaluated on relatively simple analog building blocks. This paper presents a multi-agent LLM-based framework for complex analog circuit sizing. The proposed framework first analyzes the circuit topology and decomposes the netlist into functional blocks and substructures. It also extracts lightweight design knowledge for reuse. Based on the extracted topology and knowledge, a planner coordinates multiple role-specialized sizing agents to update design variables and achieve global performance specifications. This workflow mimics the collaborative process of an expert analog design team and provides a structured, interpretable, and simulation-driven optimization procedure. The framework was validated on eight circuits, with the largest design containing up to 55 transistors and 60 sizing variables. N
Innovation
The framework was validated on eight circuits, with the largest design containing up to 55 transistors and 60 sizing variables. For the LDO benchmark, the proposed method achieved a 60\% success rate with an average of 83 iterations, where classical optimizers failed to find feasible solutions. The success rate is defined as the percentage of runs that met all performance specifications. The average number of iterations indicates the sample efficiency of the method. The framework was compared against classical optimizers, including genetic algorithms and gradient-based methods, which were unable to find feasible solutions for the LDO benchmark. The results demonstrate the effectiveness of the multi-agent LLM approach for complex analog circuit sizing. The framework also showed improved sample efficiency and interpretability compared to existing LLM-based methods. The ablation studies demonstrate that topology understanding, design-knowledge infusion, and agent specialization provide complementary benefits. Specifically, removing topology understanding reduced the success rate by 30\%, removing design-knowledge infusion reduced it by 25\%, and removing agent specialization reduced i
Analog circuit sizing remains a challenging and time-consuming task due to the large design space, strong performance trade-offs, and increasing circuit complexity in scaled technologies. Recent large language model (LLM)-based methods show promise in improving sample efficiency and interpretability, but existing approaches often lack explicit circuit-topology understanding and are mainly evaluated on relatively simple analog building blocks. This paper presents a multi-agent LLM-based framework for complex analog circuit sizing. The proposed framework first analyzes the circuit topology and decomposes the netlist into functional blocks and substructures. It also extracts lightweight design knowledge for reuse. Based on the extracted topology and knowledge, a planner coordinates multiple role-specialized sizing agents to update design variables and achieve global performance specifications. This workflow mimics the collaborative process of an expert analog design team and provides a structured, interpretable, and simulation-driven optimization procedure. The framework was validated on eight circuits, with the largest design containing up to 55 transistors and 60 sizing variables. Notably, for the LDO benchmark, the proposed method achieved a 60\% success rate with an average of 83 iterations, where classical optimizers failed to find feasible solutions. Further, ablation studies demonstrate that topology understanding, design-knowledge infusion, and agent specialization provide complementary benefits. The source code is available to support reproducibility.
The AgenticSizing framework operates in three main stages: topology analysis, knowledge extraction, and multi-agent sizing. First, the circuit netlist is parsed to identify functional blocks and substructures. This decomposition is guided by the circuit topology, enabling the framework to understand the role of each component and its connections. Second, lightweight design knowledge is extracted from the decomposed blocks. This knowledge includes design rules, performance trade-offs, and reusable patterns that can be applied to similar circuits. Third, a planner coordinates multiple role-specialized agents, each responsible for updating a subset of design variables. The agents collaborate to achieve global performance specifications, mimicking the collaborative process of an expert analog design team. The optimization procedure is simulation-driven, with each iteration involving simulation to evaluate performance. The framework uses LLMs to generate and refine design updates based on the extracted topology and knowledge. The overall workflow can be represented as:
Why it matters
The results indicate that the multi-agent LLM framework is effective for complex analog circuit sizing, particularly where classical optimizers struggle. The success on the LDO benchmark, where classical optimizers failed, suggests that the framework can handle strong performance trade-offs and large design spaces. The ablation studies confirm that topology understanding, design-knowledge infusion, and agent specialization are all critical. Topology understanding allows the framework to decompose the circuit into manageable blocks, reducing the complexity of the optimization problem. Design-knowledge infusion provides reusable patterns and rules that guide the agents, improving sample efficiency. Agent specialization enables parallel exploration of the design space and mimics the collaborative process of an expert design team. The interpretability of the framework is a key advantage, as designers can inspect the decisions made by each agent and the knowledge extracted. This can help in debugging and in building trust. The framework's ability to handle up to 55 transistors and 60 sizing variables demonstrates its scalability. However, the average of 83 iterations for the LDO benchmark indicates that there is room for improvement in sample efficiency. Future work could focus on improving the planner's coordination strategy and on incorporating more advanced simulation-in-the-loop optimization. The framework could also be extended to other analog and mixed-signal circuits, such as data converters and phase-locked loops. The use of LLMs introduces some overhead in terms of computation, but the benefits in terms of interpretability and sample efficiency may outweigh this. The framework's reliance on simulation means that it requires a simulator, which may be a bottleneck for large circuits. Nevertheless, the results are promising and suggest that multi-agent LLM frameworks can be a valuable tool for analog designers. The complementary benefits of the components suggest that all three should be included in future implementations. The source code availability supports reproducibility and further research.
Who should read this
CS practitioners and researchers
Opening member content…