Spicing up Genetic Netlist Generation with LLMs

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

Source-linked AI summary

Spicing up Genetic Netlist Generation with LLMs

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

arXiv:2608.23317v1Aug 24, 2026cs.NEcs.ARcs.LG

TL;DR

Analog topology synthesis is difficult because small structural changes can produce nonlinear behavior, while genetic search may be costly and converge prematurely. LLM-SPICEMixer embeds IGEL, an LLM proposal operator, in a SPICE-evaluated genetic loop, and improves Iris discriminant-function synthesis relative to genetic search without LLM guidance.

Problem
Analog topology synthesis requires searching discrete structures coupled with sizing, while genetic methods can require many SPICE evaluations and converge prematurely.
Method
LLM-SPICEMixer uses IGEL to generate candidate netlists from elite circuits inside genetic search, with SPICE providing evaluation and selection alongside conventional operators.
Results
LLM-SPICEMixer improves search performance over SPICEMixer on the Iris analog discriminant-function synthesis task.
Takeaways & Limitations
For non-standard analog synthesis tasks, LLMs are most effective as proposal generators embedded within simulation-driven evolutionary search rather than standalone designers.
Takeaways & Limitations
The gap between training and validation/test reward suggests that reward design or explicit regularization may need improvement for generalization.

Abstract

from arXiv · show

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.

1 Introduction

Analog topology synthesis is difficult because discrete structural choices interact nonlinearly with sizing, while genetic search can be expensive and converge prematurely. LLM-SPICEMixer addresses this by using IGEL to propose structured netlist variations within a SPICE-evaluated evolutionary loop, and evaluates it on Iris discriminant-function synthesis.

Topology synthesis remains difficult because structural choices are discrete and strongly coupled with transistor sizing.
Purely genetic search can require many SPICE evaluations and may converge prematurely, while standalone LLM circuit generation is unreliable for non-standard tasks.
LLM-SPICEMixer extends SPICEMixer with IGEL, which proposes candidate netlists from high-quality elite circuits during genetic search.
IGEL proposals are evaluated by SPICE and selected with the same reward mechanism as mutation, crossover, and pruning.
The benchmark synthesizes transistor-level circuits whose input voltages encode four Iris features and whose output voltages represent three class scores.
The paper studies prompting strategies, decoding settings, model choices, and operator mixtures, reporting improved search performance when IGEL complements genetic operators.

2 Related Work

Prior work spans fixed-topology optimization, evolutionary topology synthesis, general circuit generation, and LLM-guided evolutionary search. LLM-SPICEMixer transfers the proposal-and-selection paradigm to transistor-level analog synthesis, where SPICE evaluates physical circuit quality.

Fixed-topology analog design has used Bayesian optimization, reinforcement learning, and LLM-based optimization, while topology synthesis remains more challenging.
Recent topology-synthesis methods often target specific circuit families, although graph-based, generative, and LLM-based approaches address more general circuit generation.
LLM-SPICEMixer builds on SPICEMixer and embeds IGEL netlist proposals in an evolutionary loop, with SPICE determining quality under task-specific testbenches and process corners.
Unlike SpiceFuzz, which uses LLM-generated SPICE netlists for simulator fuzzing, LLM-SPICEMixer targets reward-driven functional synthesis.
The paper identifies LLM-SPICEMixer as the first method, to the authors’ knowledge, to embed LLM-generated netlist proposals into evolutionary analog circuit synthesis.

3 LLM-SPICEMixer

LLM-SPICEMixer evolves SPICE netlists directly and adds IGEL as an inspiration-guided proposal operator. Elite circuits are presented to an LLM, whose candidates are simulated and integrated with conventional genetic operators.

3.1 Recap of SPICEMixer: SPICEMixer treats the SPICE netlist itself as the genome in a genetic algorithm, supporting arbitrary components and process design kits.
3.1 Recap of SPICEMixer: Crossover, mutation, and pruning generate candidates by recombining elite netlists, introducing structural variation, and merging compatible components.
3.1 Recap of SPICEMixer: SPICE maps each candidate’s task-specific simulation performance to a scalar reward, while rank-based roulette-wheel sampling selects parents from the elite set.
3.2 Inspiration-Guided Evolution with LLMs: Direct LLM circuit generation is often ineffective for uncommon circuit families because models may reproduce canonical solutions instead of discovering novel topologies.
3.2 Inspiration-Guided Evolution with LLMs: IGEL uses the LLM as a proposal operator inside genetic search and combines it with crossover, mutation, and pruning to reduce generation collapse and enable refinement across operators.
3.2 Inspiration-Guided Evolution with LLMs: At each IGEL step, three elite netlists and their rewards are prompted to the LLM, which analyzes them and outputs a candidate netlist for SPICE evaluation.
3.2 Inspiration-Guided Evolution with LLMs: Preprocessing removes topology duplicates and converts PDK-specific transistor instances into simplified MOS representations with clearer component and net names.
3.2 Inspiration-Guided Evolution with LLMs: Prompt studies vary complete-netlist versus edit-based generation and reasoning versus direct output, with reasoning and structured edits identified as strong candidates.

4 Analog Discriminant-Function Synthesis

The paper evaluates LLM-SPICEMixer by synthesizing a transistor-level analog discriminant function for Iris classification, a non-standard task without a known circuit template. The benchmark combines normalized feature-to-voltage inputs, multi-corner SPICE evaluation, and reward penalties for invalid or oversized circuits.

Task: The benchmark synthesizes an analog circuit that directly implements the Iris classification discriminant function rather than reproducing a known circuit class.IGEL must infer useful patterns from elite netlist inspirations because suitable topologies are not known a priori.
Task: The circuit uses four feature inputs and three class-score outputs, with normalized features mapped to voltages in the range [0V, 1.8V].The four inputs represent sepal length, sepal width, petal length, and petal width; the outputs represent setosa, versicolor, and virginica.
Dataset and evaluation: The standard Iris dataset is split stratifiably into 90 training, 30 validation, and 30 test samples, reserving test data for final evaluation.Training reward drives search, validation selects the best discovered circuits, and test data remains held out.
Dataset and evaluation: Each candidate is evaluated with Ngspice across 17 process, voltage, and temperature corners, and results are averaged to encourage operating-condition robustness.The set includes one nominal tt corner and 16 extreme combinations of process, supply voltage, and temperature.
Reward: The reward combines classification accuracy with output-voltage separation, encouraging correct outputs toward 1.8V and incorrect outputs toward ground.The separation term distinguishes circuits with equal accuracy but different output confidence.
Reward: Penalized rewards subtract 0.05 per violated validity check and 0.0025 × N_transistors to favor structurally valid, compact circuits.Validity checks include required outputs, connected transistor bulks, non-floating nets, and restricted terminal connections.
Dataset and evaluation: Search fitness uses averaged training reward, while averaged validation reward later selects the circuit for final testing, reducing overfitting to training data.Simulation failures or unparsable output traces affect candidate evaluation.

5 Results

LLM-SPICEMixer combines genetic operators with IGEL, an LLM proposal operator that uses elite circuits as inspirations while SPICE evaluates and selects candidates. Across evolutionary comparisons and ablations, the hybrid improves search outcomes, with performance depending on prompt design, decoding, model choice, and operator mixture.

Method: IGEL samples three elite circuits, prompts Qwen3.5 27B, and submits each generated netlist to SPICE alongside conventional operators.IGEL is invoked at half the frequency of crossover, mutation, and pruning, reducing LLM inference cost.
Comparison with Evolutionary Baselines: 0.810 median reward is achieved by LLM-SPICEMixer, versus 0.747 for SPICEMixer and 0.587 for GraCo ES.The best overall circuit reaches 0.855, compared with 0.768 for the best circuit found by SPICEMixer.
Comparison with Evolutionary Baselines: 0.063 median improvement over SPICEMixer is statistically significant with a one-sided unpaired permutation test p-value of 0.0052.The test uses results from nine independent runs.
Comparison with Evolutionary Baselines: 0.807 median test reward is obtained after selecting checkpoints by validation reward, which is 0.065 higher than SPICEMixer.The validation-selected circuit is evaluated on the held-out test split.
Best Synthesized Circuit: The best validation circuit uses only petal length and petal width and achieves 93.4% train, 88.4% validation, and 85.9% test accuracy.Its two-input structure matches prior analyses identifying these attributes as dominant contributors to Iris classification.
Ablation Studies: Prompt design affects reward: reasoning and alternating raw/diff formats perform best, while balanced decoding with T=0.7 and p_top=0.9 is strongest among tested settings.Qwen3.5 27B achieves the highest median and maximum reward, improving median reward by +0.05 over Gemma3 27B; IGEL-only search underperforms the operator mixture.

6 Conclusions and Outlook

LLM-SPICEMixer embeds LLM-generated topology proposals within simulation-driven evolutionary search, improving search performance on Iris discriminant-function synthesis while leaving SPICE as the evaluation authority.

LLM-SPICEMixer augments SPICEMixer with IGEL, generating candidate netlists from elite solutions rather than designing circuits from scratch.The proposed netlists are evaluated by SPICE within the genetic search loop.
IGEL improves search performance over SPICEMixer, reduces premature convergence, and yields compact transistor-level circuits with strong performance on the Iris classification task.
Prompt design and model choice are important in this setting, whereas decoding configuration has a comparatively smaller effect.
Outlook: The training–validation/test reward gap motivates improved reward design or explicit regularization to enhance generalization.
Outlook: The authors suggest incorporating noisy or perturbed inputs during synthesis and exploring richer LLM proposal, refinement, and repair strategies.
Overall, LLMs appear most effective for non-standard analog synthesis as proposal generators embedded in simulation-driven evolutionary search, rather than standalone designers.

A Examples of LLM Prompts and Outputs

The examples show IGEL prompts paired with LLM outputs, illustrating how alternative prompt styles formulate requests and induce netlist modifications.

Figures 4 and 5 present example IGEL prompts and corresponding LLM outputs for different prompt styles.The examples illustrate both prompt formulation and the netlist modifications proposed by the model.

B Netlists of the Best Validation Circuits

This section presents the highest-validation-reward circuit and additional strong solutions, showing their rewards, compactness, robustness information, and structural diversity.

The highest-validation-reward circuit achieves averaged train/validation/test rewards of 0.855/0.811/0.780 across 17 process, voltage, and temperature corners.Per-corner accuracies are also provided to characterize robustness across operating conditions.
Additional circuits with strong validation performance illustrate diversity among solutions discovered by LLM-SPICEMixer.
The figures include complete transistor-level netlists and examples of LLM prompts, reasoning, and generated outputs for raw and diff-style templates.
One example starts from a base netlist with reward 0.767 and inspiration netlists with rewards 0.765 and 0.768, where differences include duplicating transistor M8.
Another proposed modification duplicates transistor M7 to introduce redundancy, with the stated goal of potentially improving classifier robustness and accuracy.

C Output Waveforms Across Shuffles and Corners

The waveform evaluation examines the best validation circuits across shuffled test splits and operating corners to show how their voltage behavior varies under these conditions.

Figure 9 overlays input and output waveforms for the two best-validation circuits across all considered corners and three shuffled test splits.The plots provide insight into variation in circuit behavior across operating conditions.

D Robustness to Input-Voltage Perturbations

LLM-SPICEMixer circuits remain competitive with ML baselines without input noise and show a similar accuracy degradation trend as Gaussian voltage perturbations increase. Robustness is reported at the nominal tt corner and across all process, voltage, and temperature corners.

The synthesized circuits are competitive with ML baselines in the noiseless setting.
The robustness evaluation includes both nominal tt-corner performance and average performance across all process, voltage, and temperature corners.Table 3 reports mean, standard deviation, and median test accuracies for each noise level.
Their accuracy degrades similarly to the baselines as Gaussian input-voltage noise increases.Evaluations use noise levels σ_noise ∈ {0, 0.1, 0.2, 0.3, 0.4, 0.5}, aggregated over 16 independent realizations per noisy setting.

E Results of Ablation Studies

The ablation studies vary prompting, decoding, model choice, and operator-mixture settings while otherwise holding the default IGEL configuration fixed. The operator-mixture comparison separates proposal-step budget from LLM-call budget.

Ablations examine prompting strategy, decoding settings, model choice, and the role of the operator mixture.Final best training reward is summarized over nine independent runs using average, standard deviation, minimum, median, and maximum.
Each ablation changes one factor while keeping all other settings identical to the default IGEL configuration.
IGEL-only search is compared with the full operator mixture at matched proposal steps and at the full proposal budget.The IGEL-only setting uses 18,816 proposal steps, while the full operator mixture is reported at 18,816 and 131,072 steps.

F LLM Response Length Statistics

The paper reports response-length statistics for four evaluated LLMs, with separate reasoning-token and final-output lengths for Qwen3.5 models. These measurements provide context for the model comparison.

Response lengths are summarized across the four evaluated models using both character and word counts.
For Qwen3.5 models, thinking tokens and final output are reported separately.The table represents these values as <think> plus final output because the models produce explicit thinking tokens.