HeaRT: A Hierarchical Circuit Reasoning Tree-Based Agentic Framework for AMS Design Optimization

Souradip Poddar, Chia-Tung Ho, Ziming Wei, Weidong Cao, Haoxing Ren, David Z. PanarXiv 2025 · DOI · arXiv

Introduction

Analog and Mixed-Signal (AMS) circuit design remains one of the most challenging frontiers in Electronic Design Automation (EDA). Unlike digital design, which benefits from high levels of abstraction and standardized synthesis flows, analog design is often characterized as a "black art." It involves navigating complex, multi-dimensional trade-offs where small changes in one component can have cascading effects on the entire system's performance.

Traditional automation approaches, such as Bayesian Optimization (BO) or Reinforcement Learning (RL), typically treat circuits as "black boxes." While effective in narrow scopes, these methods suffer from high sample inefficiency, a lack of explainability, and an inability to adapt to slight changes in design specifications without restarting the entire optimization process. Recent attempts to leverage Large Language Models (LLMs) for circuit design have faced hurdles like high hallucination rates and a lack of "architectural awareness"—the models often fail to understand the functional hierarchy of a complex netlist.

To address these challenges, we present HeaRT (Hierarchical Circuit Reasoning Tree), an agentic framework designed to bring human-like intuition and hierarchical reasoning to AMS design. HeaRT decomposes complex circuits into functional sub-blocks, allowing LLMs to perform targeted, context-aware optimizations.

The HeaRT Framework

The HeaRT framework operates in two distinct phases: an offline construction phase and an online agentic retrieval phase.

Phase 1: Analytical-Guided Reasoning Tree Construction (Offline)

The core of HeaRT is the transformation of a flat SPICE netlist into a structured, hierarchical knowledge graph. This process begins with Graph-Guided Decomposition. By applying Kirchhoff’s Current Law (KCL) and graph theory, the system traces DC current paths to identify fundamental functional units, such as current mirrors, differential pairs, and active loads.

Following decomposition, a Bottom-Up Consolidation step uses an LLM to analyze these identified sub-circuits. The LLM assigns semantic meaning to each block, building a "Hierarchical Circuit Reasoning Tree." This tree captures the "intent" of the design—for instance, recognizing that a group of transistors forms a comparator, which is part of a larger SAR ADC. This hierarchical abstraction mirrors how human experts approach design, significantly reducing the cognitive load on the LLM during optimization.

Overview of the HeaRT framework, illustrating the offline construction of the Hierarchical Circuit Reasoning Tree and the online agentic retrieval process for downstream optimization tasks.

Phase 2: Agentic Retrieval (Online)

When a designer initiates a task—such as optimizing a specific performance metric—the HeaRT agent traverses the reasoning tree to identify the "dominant contributor" sub-circuits. Instead of a global search, the agent performs Agentic Retrieval from a "Topology-Keyed Knowledge Database." This database contains silicon-proven circuit structures ranked by performance. By "plugging in" a superior sub-block topology and refining its sizing, HeaRT achieves targeted improvements while leaving the rest of the validated system intact.

Experimental Evaluation

HeaRT was evaluated against a benchmark of 40 AMS designs, categorized by complexity:

  • Simple (<20 devices): Basic operational amplifiers and mirrors.
  • Medium (20–40 devices): Bandgap voltage references and regulators.
  • Hard (>40 devices): Complex systems like 8-bit SAR ADCs and Analog Front-Ends (AFEs).

The framework was compared against several baselines, including ADO-LLM, AnaFlow, and vanilla frontier models like GPT-4o.

Key Results

The experimental results demonstrate significant advantages in reasoning accuracy, efficiency, and optimization speed:

  • Reasoning Accuracy: HeaRT achieved over 97% accuracy in identifying sub-circuit functions and over 98% Pass@1 performance, maintaining stability even as circuit complexity increased.
  • Token Efficiency: By focusing on specific branches of the reasoning tree, HeaRT operates at less than 0.5x the real-time token budget of previous LLM-based methods.
  • Optimization Convergence: HeaRT reached target performance metrics approximately 3x faster than standalone black-box optimizers.

A critical innovation introduced in this work is the Prior Circuit Knowledge Retention Index (PCKRI). This metric quantifies how much of the original design's structure and parameters are preserved during an update. As shown in the performance analysis, HeaRT's "scoped sizing" approach maintains a significantly higher PCKRI compared to full-circuit re-optimization, which often destroys the original design intent.

Comparison of optimization performance (Figure of Merit) and design intent preservation (PCKRI) between full sizing optimization and HeaRT's scoped sizing approach.

Conclusion

HeaRT represents a significant step toward truly intelligent EDA tools. By combining analytical graph-based decomposition with the reasoning capabilities of LLMs, it provides a transparent, traceable, and highly efficient path for AMS design optimization. Future work will focus on automating the topology database ranking through simulation metadata and extending the framework to advanced process nodes.

Summary updated

Abstract

Conventional AI-driven AMS design automation algorithms remain constrained by their reliance on high-quality datasets to capture underlying circuit behavior, coupled with poor transferability across architectures, and a lack of adaptive mechanisms. This work proposes HeaRT, a hierarchical circuit reasoning-based agentic framework for automation loops and a step toward adaptive, human-style design optimization. HeaRT consistently improves F1(subcircuits) by >= 13.5% and F1(loops) by >= 37.8% over few-shot prompting baselines across multiple LLM backbones on our 40-circuit AMS benchmark of flattened SPICE netlists, even as circuit complexity increases. Our experiments further show that HeaRT achieves >= 3x faster convergence in incremental design adaptation tasks under specification shifts across diverse optimization approaches, supporting both topology reconfiguration and sizing.
BibTeX
@misc{poddar2025heart,
  author = {Souradip Poddar and Chia-Tung Ho and Ziming Wei and Weidong Cao and Haoxing Ren and David Z. Pan},
  title = {{HeaRT: A Hierarchical Circuit Reasoning Tree-Based Agentic Framework for AMS Design Optimization}},
  year = {2025},
  eprint = {2511.19669},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.19669},
}
Figure