Research map / Large language models / Language, safety and interpretability
LaMAGIC2: Advanced Circuit Formulations for Language Model-Based Analog Topology Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LaMAGIC2 with a succinct float-input canonical formulation with identifiers (SFCI) that fine-tunes Flan-T5-base to generate power-converter topologies from voltage/efficiency specs with O(|V|) output length.
Problem
Language-model-based analog topology generation from specifications suffers from O(|V|^2) output token length and low sensitivity to precise numeric inputs, hurting success under tight tolerances and transfer to larger circuits.
Approach
Represent circuits as sparse canonical edge sets with component-type tokens plus identifiers (e.g., Sa 0), remove natural-language descriptors from inputs, and encode duty cycle as a single token, yielding linear-length outputs and better numeric precision.
Key results
- SFCI/SFM reach 0.91 success rate at tolerance 0.01 vs 0.41 for the RL-search method on 350 specs.
- SFCI achieves 34% higher success rate at tolerance 0.01 than LaMAGIC FM and about 10X lower voltage MSE (0.0006 vs 0.0061).
- On 6-component transfer with 2000 finetuning circuits, SFCI reaches 0.84 success at tolerance 0.1 vs FM's 0.76.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SFCI instead of FM | Flan-T5-base topology generation · 3, 4, 5-component circuits | voltage MSE: 0.0061 -> 0.0006 | Helped |
| use SFCI instead of FM | Flan-T5-base topology generation · 3, 4, 5-component circuits | efficiency MSE: 0.0128 -> 0.0002 | Helped |
| use SFCI instead of FM | Flan-T5-base topology generation · 6-component circuits, 2000 training examples | voltage MSE: 0.1014 -> 0.0580 | Helped |
| use SFCI instead of FM | Flan-T5-base topology generation · 6-component circuits, 2000 training examples | efficiency MSE: 0.0865 -> 0.0418 | Helped |
| add the duty-cycle prefix (ablation) | SFCI · 3, 4, 5-component circuits | voltage MSE: 0.0033 -> 0.0006 | Helped |
| add the duty-cycle prefix (ablation) | SFCI · 3, 4, 5-component circuits | efficiency MSE: 0.0028 -> 0.0002 | Helped |
| retain component-type tokens in the output (ablation) | SFCI · 3, 4, 5-component circuits | voltage MSE: 0.0012 -> 0.0006 | Helped |
| retain component-type tokens in the output (ablation) | SFCI · 3, 4, 5-component circuits | efficiency MSE: 0.0022 -> 0.0002 | Helped |
| replace FM with SFCI | Flan-T5-base topology generation · 3, 4, 5-component circuits, tolerance 0.01 | success rate: 0.67 -> 0.91 | Helped |
| use SFCI rather than SFM | Flan-T5-base topology generation · training and inference | training steps: 8,943 -> 6,886; 23.0% reduction | Helped |
Surprising findings
- o1 with few-shot prompting performs poorly (voltage MSE 0.602), indicating SFT is needed for this task.
Limitations
- Evaluated only on power converter topologies with a fixed set of component types and five duty cycles.
- Transistor-based extension is only illustrated with an example inverter, not evaluated.
- RL baseline constrained to voltage ratio only, making comparison not fully like-for-like.
Experimental setup
Datasets: LaMAGIC power converter dataset (3,4,5-component circuits) 6-component circuits from LaMAGIC dataset
Benchmarks: LaMAGIC power converter test set (12k, 9k for 6-component)
Metrics: success rate under tolerance 0.01-0.1 MSE of voltage conversion ratio and efficiency
Baselines: LaMAGIC CF LaMAGIC PM LaMAGIC FM RL-search power converter method o1 few-shot
Builds on: LaMAGIC Flan-T5 NGSPICE
Closest papers by meaning
- LaMAGIC: Language-Model-based Topology Generation for Analog Integrated Circuits (ICML 2024 · similarity 0.96)
- AnalogToBi: Device-Level Analog Circuit Topology Generation via Bipartite Graph and Grammar Guided Decoding (NeurIPS 2026 · similarity 0.91)
- AnalogGenie: A Generative Engine for Automatic Discovery of Analog Circuit Topologies (ICLR 2025 · similarity 0.88)
- AnalogGenie-Lite: Enhancing Scalability and Precision in Circuit Topology Discovery through Lightweight Graph Modeling (ICML 2025 · similarity 0.88)
- FALCON: An ML Framework for Fully Automated Layout-Constrained Analog Circuit Design (NeurIPS 2025 · similarity 0.88)
- AUTOCIRCUIT-RL: Reinforcement Learning-Driven LLM for Automated Circuit Topology Generation (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.