Introduction
Analog circuit design is frequently characterized as "black magic" due to its reliance on complex physics, fundamental circuit laws like KCL and KVL, and deep-seated engineering intuition. While general-purpose Large Language Models (LLMs) have excelled in software engineering by leveraging vast open-source codebases, they often struggle with the niche and hierarchical reasoning required for analog design. These models typically lack exposure to the specialized knowledge needed to bridge the gap between transistor-level schematics and complex systems like Phase-Locked Loops (PLLs).
The development of AnalogSeeker addresses three primary challenges: the scarcity of high-quality open-source analog design data, the intricate coupling of circuit variables, and the risk of "catastrophic forgetting" during specialized fine-tuning.
Methodology
The researchers developed AnalogSeeker using a structured three-pronged strategy designed to inject domain expertise while preserving general reasoning capabilities.
Knowledge-Based Corpus Collection
The process began with the curation of 20 canonical textbooks, spanning circuit theory, fundamentals, integrated circuits, and advanced topics. This effort resulted in a 7.26M-token Markdown corpus.
Granular Knowledge Distillation
To transform raw text into actionable training data, the team employed a multi-agent framework. As shown in the distillation workflow, one agent generates questions from specific "learning nodes" within the textbooks, while a reasoning agent (DeepSeek-R1) produces a QTSA (Question-Thinking-Solution-Answer) response. This format emphasizes the "Thinking" component, forcing the model to learn the underlying reasoning process rather than mere memorization.
Neighborhood Self-Constrained SFT (NSC-SFT)
To prevent the model from losing its general intelligence during fine-tuning, the authors introduced NSC-SFT. This method applies a Kullback–Leibler (KL) divergence penalty to ensure the fine-tuned model's output distribution remains close to the original reference model, acting as a safety rail against model collapse.
Datasets and Training
The training utilized a 112.65M-token SFT dataset containing 15.31k labeled QTSA entries. Evaluation was conducted using AMSBench-TQA, a specialized benchmark designed by experts to assess LLM proficiency in analog and mixed-signal circuits.
A key finding during training was that "Instruct" models (e.g., Qwen2.5-32B-Instruct) served as better starting points than "reasoning-heavy" models (e.g., QwQ-32B), which exhibited fragility and performance degradation when fine-tuned on narrow technical data.
Results
AnalogSeeker achieved a state-of-the-art accuracy of 85.04% on the AMSBench-TQA benchmark. This performance represents a 15.67% improvement over its base model and surpasses major commercial models, including DeepSeek-v3 (84.41%) and GPT-4o (73.99%).
Beyond benchmarks, the model demonstrated practical utility in qualitative tests. For instance, it correctly identified complex relationships in CMOS op-amp noise reduction where standard models failed. Furthermore, when integrated into the Atelier design framework, AnalogSeeker successfully assisted in the iterative design of a three-stage operational amplifier, refining components to meet specific gain and bandwidth targets.
Conclusion
AnalogSeeker represents a significant milestone as the first open-source foundation model for analog design. By prioritizing structured reasoning (QTSA) and constrained fine-tuning (NSC-SFT), it bridges the gap between general LLM capabilities and specialized engineering expertise. Future work will focus on scaling data resources and enhancing the model's ability to perform autonomous, end-to-end circuit optimization.