AnalogSeeker: An Open-source Foundation Language Model (Aug 2025)

Title: AnalogSeeker: An Open-source Foundation Language Model for Analog Circuit Design (Aug 2025)
Link: http://arxiv.org/abs/2508.10409v1
Date: August 2025

Summary:
The paper introduces AnalogSeeker, an open-source foundation language model for analog circuit design. It addresses data scarcity by curating textbooks into a textual corpus, uses a granular domain knowledge distillation method to create a learnable dataset, and employs a fine-tuning-centric training paradigm with a neighborhood self-constrained supervised fine-tuning (NSC-SFT) algorithm. AnalogSeeker achieves 85.04% accuracy on AMSBench-TQA, showing potential for design assistance.

Key Topics:
Analog circuit design
Foundation language model
Domain knowledge distillation
Supervised fine-tuning (SFT)
Electronic design automation (EDA)
Neighborhood self-constrained SFT (NSC-SFT)

Chapters:
00:00 - Introduction to Analog Circuit Design Challenges
00:38 - Analog Seeker's Key Takeaways
01:35 - Analog Seeker's Training Algorithm: NSESFT
02:32 - Analog Seeker's Potential Applications
02:55 - The Analog Design Dilemma
03:45 - Limitations of General Purpose LLMs
04:24 - Challenges in Building a Foundational Model
05:39 - Training Difficulties
06:09 - Knowledge Acquisition
07:22 - Distilling Deep Insights
07:51 - QTSA Format Explained
08:31 - Multi-Agent Framework for QTSA Generation
09:37 - Choosing the Right LLM
10:46 - Instruct Models
11:19 - Fine Tuning and SFT
12:23 - NSC SFT
13:53 - Noise Analysis in CMS OS O-PAMP
14:27 - Asymmetric Memory Management
15:14 - Benchmarks
16:24 - Opamp Design Assistance
17:57 - Realistic Design Process Exploration
18:22 - Summary of Analog Seeker
18:55 - Future Implications