Spicing up Genetic Netlist Generation with LLMs

Stefan Uhlich; Yağız Gençer; Andrea Bonetti; Arun Venkitaraman; Chia-Yu Hsieh; Eisaku Ohbuchi; Lorenzo Servadei

cs.NEAug 24, 2026

Spicing up Genetic Netlist Generation with LLMs

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

Organizations: Sony AI, Switzerland · EPFL, Switzerland · Sony Semiconductor Solutions, Japan · TU Munich, Germany

Abstract

Analog circuit topology synthesis remains challenging because useful designs occupy a tiny fraction of a combinatorial search space, and small structural changes can induce highly nonlinear changes in behavior. Evolutionary algorithms are attractive because they can optimize over discrete circuit topologies using only black-box evaluations, but they often require many SPICE simulations and may converge prematurely. We introduce LLM-SPICEMixer, a hybrid synthesis framework that augments genetic netlist generation with IGEL (Inspiration-Guided Evolution with LLMs), an LLM-based proposal operator. During search, IGEL prompts an LLM with high-performing circuits from the elite set and instructs it to generate a new SPICE netlist, which is then evaluated by SPICE and selected using the same reward mechanism as conventional genetic operators. Thus, the LLM contributes structured topology proposals while simulation remains the source of truth. We evaluate LLM-SPICEMixer on a challenging benchmark task: synthesizing transistor-level circuits that implement a discriminant function for Iris classification. Compared with the genetic framework without LLM guidance, LLM-SPICEMixer improves the median final training reward by 8.4% and the median validation-selected test reward by 8.8%. The best validation-selected circuit achieves 93.3% test accuracy at the nominal tt corner and 85.9% average test accuracy across 17 process, voltage, and temperature corners.

Figures & tables

Figure 1. Overview of LLM-SPICEMixer: (a) LLM-augmented genetic circuit optimization loop. (b) IGEL prompts the LLM with inspiration circuits, then evaluates the proposed netlist with SPICE; it is added to the elite set if good enough.

Figure 2. Training and validation rewards of the best-so-far circuit in each run. Circuits are selected by training reward and then evaluated on validation. Each run uses $131{,}072$ proposal steps; thicker lines indicate medians over nine runs.

	GraCo ES ( 41 )	SPICEMixer ( 40 )	LLM-SPICEMixer
Average $\pm$ Std. Dev.	0.588 $\pm$ 0.022	0.737 $\pm$ 0.031	0.799 $\pm$ 0.039
Minimum	0.558	0.668	0.719
Median	0.587	0.747	0.810
Maximum	0.626	0.768	0.855

Table 1. Final best training reward for different methods. Values are computed over nine independent runs.

	GraCo ES ( 41 )	SPICEMixer ( 40 )	LLM-SPICEMixer
Average $\pm$ Std. Dev.	0.627 $\pm$ 0.025	0.730 $\pm$ 0.031	0.782 $\pm$ 0.041
Minimum	0.580	0.667	0.715
Median	0.622	0.742	0.807
Maximum	0.670	0.773	0.824

Table 2. Test rewards obtained by selecting the checkpoints with the highest validation reward. Values are computed over nine independent runs.

Panel of Figure 3 — Figure 3. Schematic of the best validation circuit found by LLM-SPICEMixer. The circuit uses only input_2 and input_3 , corresponding to petal length and petal width, respectively. The output nets output_0 , output_1 , and output_2 correspond to setosa, versicolor, and virginica.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 4 — Figure 4. LLM prompt and answer for the “raw”-style template version.

Panel of Figure 5 — Figure 5. LLM prompt and answer for the “diff”-style template

Panel of Figure 6 — Figure 6. Netlist of the circuit with the best validation reward found by LLM-SPICEMixer.

Panel of Figure 7 — Figure 7. Netlists of additional circuits with strong validation performance.

Panel of Figure 8 — Figure 8. Voltage transients evaluated on the test split. The plots overlay the waveforms obtained from the three shuffled versions of the test split.

Panel of Figure 9 — Figure 9. Voltage transients evaluated on the test split. The plots overlay the waveforms obtained for all corners and all three shuffled versions of the test split.

Panel of Figure 10 — Figure 10. Robustness of LLM-SPICEMixer compared with ML baselines under input-voltage perturbations. The synthesized circuits are competitive with the baselines in the noiseless setting and degrade similarly as Gaussian noise $\sigma_{\text{noise}}$ is added to the inputs, indicating that the analog solutions retain useful classification margins under imperfect sensor readings. Dashed lines represent chance accuracy.

Method	Corners	$\sigma_{\text{noise}}=0$	$\sigma_{\text{noise}}=0.1$	$\sigma_{\text{noise}}=0.2$	$\sigma_{\text{noise}}=0.3$	$\sigma_{\text{noise}}=0.4$	$\sigma_{\text{noise}}=0.5$
Logistic regression	N/A	$90.0\pm 0.0$ / $90.0$	$87.0\pm 2.1$ / $86.7$	$79.5\pm 3.7$ / $80.0$	$70.8\pm 4.6$ / $70.6$	$64.3\pm 5.2$ / $66.1$	$59.8\pm 5.1$ / $59.4$
Single-hidden-layer network	N/A	$90.0\pm 0.0$ / $90.0$	$89.6\pm 1.9$ / $90.0$	$79.7\pm 3.2$ / $80.0$	$69.8\pm 3.5$ / $70.6$	$63.7\pm 4.2$ / $63.9$	$57.8\pm 4.0$ / $58.3$
LLM-SPICEMixer (best validation)	tt	$93.3\pm 0.0$ / $93.3$	$89.2\pm 1.9$ / $89.4$	$84.1\pm 3.7$ / $83.9$	$76.0\pm 12.0$ / $78.9$	$65.6\pm 15.7$ / $71.1$	$55.0\pm 15.9$ / $63.3$
LLM-SPICEMixer (best validation)	all	$85.9\pm 0.0$ / $85.9$	$84.2\pm 1.3$ / $84.2$	$81.0\pm 2.6$ / $81.1$	$74.8\pm 3.9$ / $74.4$	$63.4\pm 8.5$ / $63.4$	$52.1\pm 9.6$ / $53.9$
LLM-SPICEMixer ( $2^{\mathrm{nd}}$ best validation)	tt	$93.3\pm 0.0$ / $93.3$	$90.0\pm 2.5$ / $90.0$	$83.5\pm 4.1$ / $85.0$	$78.2\pm 4.5$ / $77.8$	$72.9\pm 4.3$ / $73.3$	$60.3\pm 13.9$ / $64.4$
LLM-SPICEMixer ( $2^{\mathrm{nd}}$ best validation)	all	$87.5\pm 0.0$ / $87.5$	$86.1\pm 1.8$ / $86.3$	$82.0\pm 2.8$ / $82.4$	$76.0\pm 4.5$ / $76.5$	$67.5\pm 7.0$ / $69.7$	$57.0\pm 9.0$ / $59.4$

Appendix

Table 3. Test-set accuracy across noise levels. For LLM-SPICEMixer, both tt -only and all-corner results are shown. Each cell reports mean

\pm

standard deviation / median, in percent (

\%

	without reasoning			with reasoning
	“raw”-style	“diff”-style	alternating	“raw”-style	“diff”-style	alternating
Average $\pm$ Std. Dev.	0.731 $\pm$ 0.031	0.736 $\pm$ 0.020	0.733 $\pm$ 0.036	0.731 $\pm$ 0.038	0.745 $\pm$ 0.045	0.745 $\pm$ 0.042
Minimum	0.697	0.704	0.681	0.672	0.658	0.677
Median	0.713	0.745	0.729	0.750	0.754	0.763
Maximum	0.788	0.757	0.779	0.800	0.802	0.808

Appendix

Table 4. Ablation: final best reward for different prompting strategies on the training split. Values are computed over nine independent runs.

	Deterministic	Conservative	Balanced	Entropy	Entropy++
	$T=0$ , $p_{\text{top}}=1.0$	$T=0.3$ , $p_{\text{top}}=0.9$	$T=0.7$ , $p_{\text{top}}=0.9$	$T=1.0$ , $p_{\text{top}}=0.95$	$T=1.3$ , $p_{\text{top}}=0.98$
Average $\pm$ Std. Dev.	0.740 $\pm$ 0.028	0.734 $\pm$ 0.037	0.745 $\pm$ 0.042	0.726 $\pm$ 0.040	0.744 $\pm$ 0.040
Minimum	0.699	0.650	0.677	0.658	0.679
Median	0.741	0.756	0.763	0.748	0.755
Maximum	0.798	0.772	0.808	0.771	0.798

Appendix

Table 5. Ablation: final best reward for different decoding settings on the training split. Values are computed over nine independent runs.

	Gemma3 270M	Gemma3 1B	Gemma3 4B	Gemma3 12B	Gemma3 27B	Qwen3.5 9B	Qwen3.5 27B
Average $\pm$ Std. Dev.	0.737 $\pm$ 0.034	0.740 $\pm$ 0.039	0.746 $\pm$ 0.028	0.745 $\pm$ 0.042	0.750 $\pm$ 0.019	0.743 $\pm$ 0.040	0.799 $\pm$ 0.040
Minimum	0.686	0.671	0.701	0.677	0.709	0.671	0.719
Median	0.737	0.749	0.744	0.763	0.760	0.750	0.810
Maximum	0.801	0.790	0.793	0.808	0.777	0.812	0.855

Appendix

Table 6. Ablation: final best reward for different models on the training split. Values are computed over nine independent runs.

	All operators $18{,}816$ steps ( $16{,}128$ SPICEMixer + $2{,}688$ IGEL)	Only IGEL $18{,}816$ steps ( $0$ SPICEMixer + $18{,}816$ IGEL)	All operators $131{,}072$ steps ( $112{,}347$ SPICEMixer + $18{,}725$ IGEL)
Average $\pm$ Std. Dev.	0.705 $\pm$ 0.031	0.607 $\pm$ 0.035	0.745 $\pm$ 0.042
Minimum	0.665	0.538	0.677
Median	0.689	0.619	0.763
Maximum	0.744	0.656	0.808

Appendix

Table 7. Ablation: final best reward for Gemma3 12B when using the full operator mixture (= LLM-SPICEMixer) versus using only IGEL. The first two columns compare both settings at the same number of proposal steps. The last column reports the full all-operator run for reference; the IGEL-only setting uses approximately the same number of LLM calls as this full run.

	Gemma3 12B	Gemma3 27B	Qwen3.5 9B	Qwen3.5 27B
Number of characters
Average $\pm$ Std. Dev.	$1{,}576.7\pm 389.2$	$1{,}215.9\pm 336.0$	$5{,}401.5\pm 4{,}217.8\hskip 9.24994pt$ $\hskip 9.24994pt+\penalty\ 1{,}029.4\pm 625.6$	$21{,}825.9\pm 8{,}690.7\hskip 9.24994pt$ $\hskip 9.24994pt+\penalty\ 1{,}322.9\pm 725.3$
Minimum	430	263	$1{,}173+61$	$1{,}216+62$
Median	1,532	1,176	$3{,}961+894$	$23{,}871+1{,}336$
Maximum	3,780	16,977	$59{,}653+23{,}750$	$69{,}535+23{,}088$
Number of words

Appendix

Table 8. Response-length statistics across the four models. For the Qwen3.5 models, values are reported as <think>

+

final output, because these models produce explicit thinking tokens.

Explore similar work

AaLLM: An End-to-End Analog Circuit Design Framework from Topology Generation to Sizing Using Large Language Models
Aug 13, 2026Mohammed Ayman Habib, Rylan Hart, Morteza FayaziMulti-Agent LLM Systems Electronic Design Automation
University of Utah, Salt Lake City, Utah, USA
NetlistBench: Evaluating LLM Reliability in SPICE Netlist Recognition and Manipulation
Aug 12, 2026Jiarui Ma, Jianghan Wang, Yuheng Ma +2LLM Evaluation Benchmark Design
Southern University of Science and Technology School of Microelectronics Shenzhen, China
EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models
Jul 15, 2026Guanglei Zhou, Chen-Chia Chang, Yikang Shen +5Large Language Model-Guided Optimization Electronic Design Automation
Electrical and Computer Engineering, Duke University, Durham, USA · MIT-IBM Watson AI Lab, Cambridge, USA · IBM T. J. Watson Research Center, Yorktown Heights, USA

Aug 13, 2026J/K move · Enter open · S save

AaLLM: An End-to-End Analog Circuit Design Framework from Topology Generation to Sizing Using Large Language Models

Mohammed Ayman Habib, Rylan Hart, Morteza Fayazi

University of Utah, Salt Lake City, Utah, USA

Multi-Agent LLM Systems Electronic Design Automation

Open paper AskPDF arXiv

Analog circuit design is a time-consuming, iterative process in a nonlinear and high-dimensional design space that relies heavily on expert intuition. Among recent developments, LLMs have introduced a promising approach by bringing natural language reasoning to circuit design tasks. The majority of conventional LLM-based approaches provide fragmented solutions that focus either only on sizing or topology generation. These methods require adding specific technical knowledge manually, which is inefficient and prone to hallucinations during circuit sizing. Moreover, the inherent trade-off in meeting different specs makes current approaches iterative and tedious. Another shortcoming is the inability to create innovative topologies, which may lead to sub-optimal designs due to reliance on conventional topologies. In this paper, we present AaLLM, an open-source end-to-end multi-agent LLM workflow that takes user specs as input and outputs the appropriate netlist, encompassing both topology generation and circuit sizing. AaLLM automates the creation of a relevant knowledge base from research papers and textbooks to combat tedious manual data collection. A RAG model is implemented to emulate circuit design expertise using this knowledge base. Moreover, AaLLM uses a novel tri-agent feedback system comprising a Designer that determines circuit component values, a Critic that scrutinizes these values, and an Evaluator that minimizes circuit sizing iterations by arbitrating between the other two agents. AaLLM-generated novel topologies achieve a figure of merit (FoM) comparable to that of known topologies, and up to 3x higher for certain circuits. Testing on several circuit topologies, our results show a 3x - 4.5x decrease in the number of SPICE calls at inference when compared to SOTA multi-agent LLM pipelines. The results also show a 40x decrease in wall-clock time compared to existing approaches.

Abstract

Explore similar work

AaLLM: An End-to-End Analog Circuit Design Framework from Topology Generation to Sizing Using Large Language Models

NetlistBench: Evaluating LLM Reliability in SPICE Netlist Recognition and Manipulation

EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models