🤖 machine learning

Spicing up Genetic Netlist Generation with LLMs

The paper introduces LLM-SPICEMixer, a hybrid framework that integrates an LLM-based proposal operator (IGEL) into genetic netlist generation to enhance analog circuit topology synthesis, achieving significantly higher accuracy and robustness in a benchmark Iris classification task compared to traditional methods.

Original authors: Stefan Uhlich, Yağız Gençer, Andrea Bonetti, Arun Venkitaraman, Chia-Yu Hsieh, Eisaku Ohbuchi, Lorenzo Servadei

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Stefan Uhlich, Yağız Gençer, Andrea Bonetti, Arun Venkitaraman, Chia-Yu Hsieh, Eisaku Ohbuchi, Lorenzo Servadei

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Designing analog circuits is a bit like trying to find a needle in a haystack, where the haystack is made of billions of possible arrangements and the needle is a working device. In the world of electronics, analog circuits are the components that handle continuous signals, such as the voltage from a microphone or the current from a solar panel. Unlike digital chips that simply switch on and off, these circuits must be tuned with extreme precision; a tiny change in how a transistor is connected can cause the entire device to fail or behave unpredictably. For decades, engineers have relied on complex computer algorithms to search for these working designs, but the process is often slow and prone to getting stuck in dead ends. The challenge is finding a way to explore the vast space of possibilities without wasting time on designs that will never work.

A team of researchers from Sony AI and several universities has introduced a new approach that blends traditional computer search methods with the creative capabilities of large language models. These models, which are the same technology behind advanced chatbots, are typically trained to write code or answer questions. The researchers wondered if they could use them not to design a circuit from scratch, but to act as a creative assistant that suggests small, smart improvements to existing designs. They built a system called LLM-SPICEMixer, which runs a continuous loop of trial and error. In this loop, a computer simulation tests thousands of circuit designs, keeping the best ones and discarding the rest. The new twist is that when the system needs a fresh idea, it asks the language model to look at the best designs found so far and propose a new variation.

The researchers tested this system on a specific task: creating a transistor-level circuit that could sort data about iris flowers into three categories. This is a classic problem in machine learning, but here the goal was not to write software code, but to build a physical electronic circuit that could perform the same classification using voltage levels. The circuit had to take four input signals representing flower measurements and produce three output signals indicating the flower type. Because there is no standard blueprint for such a circuit, the system had to discover a working structure from scratch. The language model was given the top-performing circuits from the search and asked to imagine a better version, much like a human engineer might sketch a modification on a blueprint.

The results showed that adding the language model significantly improved the search process. Without the model, the traditional search method often converged too quickly, settling on a mediocre design and failing to find better options. With the model's suggestions, the system found circuits that were more accurate and robust. The best circuit discovered by the hybrid system achieved a test accuracy of 93.3 percent under ideal conditions and maintained an average accuracy of 85.9 percent across a wide range of temperatures and voltages. This performance was notably higher than what the traditional method achieved on its own. The researchers found that the language model did not simply copy existing designs; it learned to combine useful patterns from the best circuits to create new topologies that the traditional algorithm had missed.

However, the study also clarified what the language model cannot do. When the researchers asked the model to design a complete circuit from scratch without any examples, it tended to produce repetitive or non-functional designs. The model worked best when it was guided by the high-quality examples provided by the search loop. This suggests that for complex engineering tasks, artificial intelligence is most effective when used as a partner that refines and evolves human or machine-generated ideas, rather than as a standalone designer. The system also discovered that the best circuits often relied on only two of the four available flower measurements, a finding that aligns with known statistical properties of the data. This implies that the system was not just guessing, but was actually learning to identify the most important features for the task.

The researchers evaluated their findings by running the search process nine times independently to ensure the results were consistent and not just a lucky fluke. They compared their method against other advanced algorithms and found that their hybrid approach consistently produced superior results. They also tested how well the circuits held up when the input signals were slightly noisy, simulating real-world imperfections. The circuits designed with the help of the language model remained stable and accurate even when the input data was distorted, showing that the search process had found solutions that were not just mathematically correct but physically robust.

This work demonstrates a promising path forward for automated circuit design. By treating the language model as a proposal generator within a rigorous simulation loop, the researchers were able to overcome some of the limitations of traditional search methods. The system did not replace the need for precise physical simulation; instead, it used the simulation as a source of truth to validate the model's creative suggestions. The success of this approach suggests that for tasks where no standard solution exists, combining the pattern recognition of large language models with the reliability of physical simulation can lead to the discovery of novel and efficient electronic designs. The findings indicate that while artificial intelligence may not yet be ready to replace human engineers entirely, it can serve as a powerful tool to accelerate the discovery of new circuit architectures that would be difficult to find through conventional means alone.

Technical Summary: Spicing up Genetic Netlist Generation with LLMs

Problem Statement

Analog circuit topology synthesis remains a significant challenge in electronic design automation. Unlike sizing optimization for fixed topologies, which has seen progress via Bayesian optimization and reinforcement learning, topology synthesis involves discrete structural choices that are strongly coupled with sizing. The search space is combinatorial, with useful designs occupying a tiny fraction of possible configurations. Small structural changes can induce highly nonlinear and non-smooth behavioral shifts, making gradient-based methods ineffective. While evolutionary algorithms (EAs) are attractive for optimizing discrete topologies using only black-box SPICE evaluations, they often suffer from premature convergence and require a vast number of simulations. Conversely, relying solely on Large Language Models (LLMs) to generate complete circuits from scratch is unreliable, particularly for non-standard tasks where memorized templates do not exist, often leading to the reproduction of canonical solutions rather than novel discoveries.

Methodology: LLM-SPICEMixer and IGEL

The authors introduce LLM-SPICEMixer, a hybrid synthesis framework that augments the existing genetic algorithm SPICEMixer with a novel LLM-based proposal operator called IGEL (Inspiration-Guided Evolution with LLMs).

Core Architecture

The system operates within a SPICE-driven evolutionary loop. Instead of using the LLM as a standalone designer, IGEL functions as one of several proposal operators alongside traditional genetic operators (crossover, mutation, and pruning).

Inspiration Sampling: From the current elite set of high-performing circuits, $K=3$ netlists are sampled using rank-based roulette-wheel selection.
Preprocessing: To prevent the LLM from merely adjusting transistor sizes (which leads to premature convergence on a single topology), the system filters out circuits with identical topologies. Furthermore, PDK-specific transistor instances (e.g., SkyWater 130nm) are converted into a simplified, descriptive MOS representation (e.g., renaming net_supply_0 to net_vdd) to reduce confusion from obscure net names.
IGEL Prompting: The sampled netlists and their reward scores are fed into the LLM. The prompt strategy varies between:
- "Raw" vs. "Diff": "Raw" prompts ask for a complete new netlist, while "Diff" prompts ask for specific line-by-line edits to a base netlist.
- **"With Reasoning" vs. "Without Reasoning":" "With reasoning" prompts require the LLM to first analyze the inspirations and propose a modification strategy before generating the netlist.
Evaluation: The LLM generates a candidate netlist, which is converted back to the PDK format, simulated via SPICE under multiple process, voltage, and temperature (PVT) corners, and evaluated using a reward function. If the candidate performs well, it is added to the elite set.

Reward Function

The framework optimizes for an analog discriminant function for the Iris classification task. The reward $R$ is a composite metric:
$R = \frac{1}{2}(A + M)$
Where $A$ is classification accuracy and $M$ is a target-voltage score encouraging large output separation (high voltage for the correct class, low for others). The reward is further penalized for structural invalidity (e.g., floating nets) and circuit size (number of transistors) to favor compact, physically plausible designs.

Key Contributions

Hybrid Framework: The introduction of LLM-SPICEMixer, which integrates LLMs as a proposal mechanism within a genetic search loop rather than as a standalone generator. This leverages the LLM's ability to propose structured topological variations while relying on SPICE simulation as the source of truth.
IGEL Operator: A specific operator that uses elite netlists as "inspirations" to guide the LLM, preventing the model from generating random or canonical circuits and instead encouraging the evolution of novel topologies.
Benchmark Task: The proposal of an analog discriminant-function synthesis benchmark based on the Iris dataset. This task is distinct from standard analog design (like op-amp sizing) because it requires discovering a circuit topology from scratch to implement a data-dependent classification function, with no known canonical solution.
Empirical Analysis: A comprehensive study of prompting strategies, decoding settings, model choices, and operator mixtures, demonstrating that explicit reasoning and alternating prompt styles yield the best results.

Results

The method was evaluated on the Iris classification task using 17 PVT corners.

Performance Gains: Compared to the baseline SPICEMixer (without LLM guidance), LLM-SPICEMixer improved the median final training reward by 8.4% and the median validation-selected test reward by 8.8%.
Statistical Significance: A permutation test confirmed the improvement is statistically significant ( $p=0.0052$ ).
Best Circuit: The best validation-selected circuit achieved 93.3% test accuracy at the nominal $tt$ corner and 85.9% average test accuracy across all 17 corners.
Circuit Characteristics: The best synthesized circuit was compact, utilizing only two of the four available inputs (petal length and width), implicitly discovering the dominant features of the Iris dataset. It demonstrated robustness against input voltage noise comparable to standard machine learning baselines (logistic regression and MLP).
Ablation Insights:
- Prompting: Explicit reasoning in prompts significantly improved performance. An alternating scheme between "raw" and "diff" formats provided the best diversity.
- Model Choice: Larger models performed better, with Qwen3.5 27B outperforming Gemma3 variants.
- Operator Mixture: IGEL is most effective as a complement to traditional genetic operators. An "IGEL-only" search performed substantially worse than the mixed approach, indicating that LLMs provide diverse structural proposals while genetic operators (crossover, mutation) are essential for local refinement and maintaining search dynamics.

Significance and Claims

The paper claims that for non-standard analog synthesis tasks where no canonical topology exists, LLMs are most effective not as autonomous designers but as proposal generators embedded within a simulation-driven evolutionary loop. By augmenting genetic algorithms with LLM-based proposals, the method mitigates premature convergence and discovers diverse, high-performing transistor-level circuits that traditional genetic search alone might miss. The work demonstrates that LLMs can successfully navigate the complex, discrete search space of analog topology synthesis when guided by high-quality inspirations and constrained by rigorous physical simulation.