---
title: 'AnalogSeeker: Analog Circuit Design LLM'
url: https://www.emergentmind.com/topics/analogseeker
type: topic
---

# AnalogSeeker: Analog Circuit Design LLM

AnalogSeeker is an open-source foundation language model developed specifically to address the unique data scarcity, knowledge complexity, and automation requirements of analog circuit design. Built atop a large-scale, high-quality textual corpus curated from canonical analog circuit textbooks, AnalogSeeker employs a multi-agent granular knowledge distillation method and a principled fine-tuning-centric training paradigm, introducing innovations in both training methodology and dataset construction. The model achieves state-of-the-art accuracy on dedicated analog knowledge benchmarks and demonstrates practical downstream utility in complex analog design tasks. AnalogSeeker is freely available for research at https://huggingface.co/analogllm/analogseeker [2508.10409].

## 1. Domain-Specific Corpus Collection

AnalogSeeker’s core is rooted in a systematically assembled “textual domain corpus” that comprehensively represents the analog circuit body of knowledge. The framework for corpus collection is explicitly structured into four ascending stages:

- **Circuit theory:** Covers foundational passive network laws, network analysis, and both time and frequency domain analysis.
- **Analog circuit basis:** Encompasses device characteristics (e.g., MOSFET, BJT), fundamental amplifiers, feedback theory, and stability analysis.
- **Analog integrated circuits:** Focuses on concrete modules such as operational amplifiers, comparators, and integrated CMOS design practices.
- **Advanced circuit topics:** Addresses specialized modules such as phase-locked loops (PLLs).

Twenty canonical textbooks spanning at least a dozen key analog circuit types were curated and segmented both by learning stage and circuit class. A commercial OCR and structure extraction pipeline (Mathpix) yielded a Markdown corpus of 7.26 million tokens, hierarchically structured by chapter and subsection, with explicit extraction and anonymization of mathematical expressions and figures.

## 2. Granular Knowledge Distillation via Multi-Agent Framework

To translate the dense, multifaceted textbook information into a machine-learnable supervision source, AnalogSeeker introduces a “learning node”–driven decomposition. Each subsection of the corpus (typically ~2000 tokens; 2,698 nodes in total) is transformed by a multi-agent framework into datasets of question–answer pairs, explicitly encoding reasoning steps.

The QTSA (Question–Thinking–Solution–Answer) format is employed:
- **Q(i):** Agent Ψ_Q generates an exam-style question directly from the learning node.
- **T(i) and S(i):** Agent Ψ_A, given the node and Q(i), outputs a detailed reasoning process (<think>…</think>), sequential solution steps (<solution>…</solution>), and
- **A(i):** The answer (<answer>…</answer>).

Each node is sampled Nₛ = 5 times, and standardized by a post-processing agent Ψ_P, ensuring robustness to formatting and over-specific references. This yields a fine-grained, explicit, and high-quality supervised training dataset (112.65M tokens).

## 3. Model Architecture and Training Paradigm

AnalogSeeker fine-tunes the Qwen2.5-32B-Instruct LLM, adopting a training strategy informed by both corpus size and domain complexity:

- **Fine-tuning–centric paradigm:** Given the modest size of textbook-derived unsupervised corpus, the approach eschews continual pre-training (CPT) in favor of focused supervised fine-tuning (SFT) on the distilled QTSA data. Experiments show <1% absolute improvement from CPT+SFT versus SFT alone, highlighting that classical pre-training is not cost-effective at this corpus scale.
- **Instruct Model Preference:** “Instruct” LLMs (e.g., Qwen2.5-32B-Instruct) proved more robust to further fine-tuning than “reasoning models,” with reward-optimized parameters exhibiting fragility during domain adaptation.
- **Empirical and theoretical validation:** Ablations confirm that SFT is the main contributor to analog circuit knowledge transfer, and that instruct models balance adaptability and capacity retention.

## 4. Neighborhood Self-Constrained Supervised Fine-Tuning (NSC-SFT)

To ensure effective domain adaptation without catastrophic forgetting of foundational LLM capabilities, AnalogSeeker employs NSC-SFT, which regularizes the fine-tuning trajectory:

- **Loss formulation:**
  $$
  L = L_{CE}(y_{\text{predict}}, y_{\text{label}}) + \lambda \cdot D_{KL}(p_{\text{predict}} \,\|\, p_{\text{ref}})
  $$
  - $L_{CE}$: cross-entropy loss for standard supervised learning.
  - $D_{KL}(p_{\text{predict}}\|\ p_{\text{ref}})$: Kullback–Leibler divergence between the current (fine-tuned) output distribution and the reference (pre-trained) model, over the entire output vocabulary $V$, at each input $x$:
    $$
    D_{KL}(p_{\theta}(v|x)\|\ p_{\theta_0}(v|x)) = \sum_{v \in V} p_{\theta}(v|x) \cdot \log\frac{p_{\theta}(v|x)}{p_{\theta_0}(v|x)}
    $$
  - $\lambda$ is a tunable hyperparameter.

- **Engineering realization:** Memory-peak optimization is required for 32B models with long contexts (8192 tokens), accomplished by asymmetric memory management (reference model resident on every GPU, target model distributed with DeepSpeed ZeRO-3) and careful tensor deletion.

- **Convergence guarantee:** Standard analysis of smooth composite losses guarantees convergence to stationary points for learning rates in $(0,1/L)$, $L$ being the Lipschitz constant.

## 5. Performance and Evaluation Benchmarks

AnalogSeeker establishes state-of-the-art performance on AMSBench-TQA, a benchmark designed for textual QA in analog circuit knowledge:
- **Accuracy:** 85.04%, an absolute improvement of 15.67 percentage points over the Qwen2.5-32B-Instruct baseline and outperforming both reasoning LLMs (QwQ-32B, 81.54%) and commercial models (GPT-4o, 73.99%; DeepSeek-v3, 84.41%).

On downstream operational amplifier design within the Atelier framework:
- **Capabilities:** Iterative topology design, topology modification (e.g., introducing series nulling resistors to improve phase margin), and expert-level circuit analysis in natural language.
- **Trajectory documentation:** Full design trajectories detail the agent’s reasoning and decision-making steps, including handling of phase margin and output swing.

## 6. Open Research Resource and Impact

AnalogSeeker is open-sourced for research use on HuggingFace (https://huggingface.co/analogllm/analogseeker). This public availability:
- **Facilitates reproducibility** by providing access to model weights, QTSA dataset construction protocol, and training recipes.
- **Enables domain adaptation and integration,** e.g., fine-tuning for particular analog subfields or integration with frameworks such as Atelier.
- **Accelerates research in analog EDA automation,** especially in data-constrained domains where classical LLMs lack sufficient prior.

This approach represents a marked advance in the methodological rigor of analog circuit LLM development, coupling fine-grained textbook data, explicit multi-agent knowledge distillation, purposefully constrained fine-tuning, and benchmarked validation.

## 7. Summary Table: AnalogSeeker Foundation Model

| Feature                        | Description                                                                   | Quantitative Result/Value         |
|------------------------------- |-------------------------------------------------------------------------------|-----------------------------------|
| Corpus size                    | Curated textbook corpus; Markdown, 7.26M tokens                               | 20 books, 2,698 learning nodes    |
| Knowledge distillation format  | Multi-agent QTSA (Q–Think–Solution–Answer)                                    | 112.65M tokens distilled SFT data |
| Model base                     | Qwen2.5-32B-Instruct                                                          | 32B, 8192-token context           |
| Fine-tuning method             | NSC-SFT (adds KL divergence constraint)                                       | $L = L_{CE} + \lambda D_{KL}$     |
| AMSBench-TQA accuracy          | Analog circuit QA benchmark                                                   | 85.04%                            |
| Improvement over baseline      | Compared to Qwen2.5-32B-Instruct and commercial models                        | +15.67, +0.63, and +11.05 points  |
| Availability                   | Open-source release on HuggingFace                                            | https://huggingface.co/analogllm/analogseeker |

The AnalogSeeker initiative provides both a rigorous engineering exemplar and a practical, open tool for the analog circuit design research community, demonstrating that domain-specific LLMs can close performance gaps in fields with structural data scarcity and compositional complexity.

Source: https://www.emergentmind.com/topics/analogseeker