PaperlayerView exact revision on arXiv

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:2608.23317v1cs.NEcs.ARcs.LG

TL;DR

Analog circuit topology synthesis is difficult because small structural changes can produce highly nonlinear behavior, and genetic search may require many SPICE evaluations or converge prematurely. LLM-SPICEMixer adds LLM-generated proposals from elite circuits to a simulation-driven genetic search, and improves search performance over SPICEMixer on Iris analog discriminant-function synthesis.

  • Problem

    Analog topology synthesis is challenging because small structural changes can produce highly nonlinear behavior, while genetic search may require many SPICE evaluations and converge prematurely.

    Source: §1 · p. 1

  • Method

    LLM-SPICEMixer embeds IGEL in genetic search, prompting an LLM with elite circuits to propose candidate netlists that SPICE evaluates alongside conventional genetic operators.

    Source: §1 · p. 1, §3.2 · p. 3

  • Results

    On Iris analog discriminant-function synthesis, IGEL improves search performance over SPICEMixer and reduces premature convergence.

    Source: §1 · p. 2, §6 · p. 6

  • Takeaways & Limitations

    For non-standard analog synthesis tasks, the supported conclusion is that LLMs are most effective as proposal generators within simulation-driven evolutionary search, rather than as standalone designers.

    Source: §6 · p. 7

    The training-to-validation/test reward gap indicates that generalization remains a limitation, with improved reward design or explicit regularization suggested as future work.

    Source: §6 · p. 7

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

LLM-SPICEMixer integrates elite-inspired LLM netlist proposals into SPICE-evaluated genetic search, targeting open-ended analog topology synthesis and testing it on a transistor-level Iris classification benchmark.

1 Introduction

  • Topology synthesis remains difficult because it involves discrete structural choices coupled with sizing, while small circuit changes can produce highly nonlinear, non-smooth behavioral changes.Prior methods have made progress on fixed-topology sizing, while many successful topology approaches focus on established circuit families.Source: §1 · p. 1
  • IGEL prompts an LLM with elite circuits to propose netlists, which SPICE evaluates and selects using the same reward mechanism as conventional genetic operators.The LLM adds structured topology variations to the evolutionary loop rather than replacing genetic search.Source: §1 · p. 1
  • The Iris benchmark requires discovering transistor-level structures whose inputs encode four features and outputs represent three class scores, without a known suitable topology.It targets data-dependent classification behavior rather than standard analog design targets.Source: §1 · pp. 1–2
  • The paper studies prompting strategies, decoding settings, model choices, and operator mixtures, reporting that IGEL improves search performance as a complement to genetic operators.The benchmark is intended to evaluate open-ended transistor-level topology synthesis.Source: §1 · p. 2

2 Related Work

Prior analog automation includes fixed-topology optimization, specialized topology search, and general circuit-generation methods; LLM-SPICEMixer instead uses an LLM proposal operator within genetic evolution. The paper positions it as the first evolutionary-loop method for analog synthesis to embed LLM-generated netlist proposals.

2 Related Work

  • Prior analog automation optimizes fixed topologies, while topology synthesis spans evolutionary search, circuit-family methods [21], and general generators [22] [25] [27].Agentic LLM frameworks such as AnalogAgent [28] use multi-agent workflows for iterative analog circuit generation; these methods generally use learned models or LLMs as primary generators or design agents.Source: §2 · p. 2
  • LLM-SPICEMixer extends SPICEMixer [26] with IGEL proposals from elite netlists, embedding LLM-generated netlists in an evolutionary loop for analog circuit synthesis.The paper describes it as the first method to embed LLM-generated netlist proposals into an evolutionary loop for analog circuit synthesis.Source: §2 · p. 2
  • FunSearch and related systems use LLM proposals in iterative evaluation loops, whereas SpiceFuzz generates SPICE netlists for simulator fuzzing rather than reward-driven functional synthesis.Unlike general program-search systems, LLM-SPICEMixer candidates are physical circuits evaluated by SPICE under task-specific testbenches and process corners.Source: §2 · p. 2

3 LLM-SPICEMixer

LLM-SPICEMixer combines genetic evolution of SPICE netlists with an LLM proposal operator, while SPICE simulation evaluates candidate circuits. The section describes the search, preprocessing, and prompt-design choices that support this approach.

3.1 Recap of SPICEMixer

  • SPICEMixer evolves SPICE netlists directly as genomes, using crossover, mutation, and pruning to generate candidates evaluated by SPICE and selected through a reward-based elite set.Parents are sampled from the elite set by rank-based roulette-wheel selection, balancing preference for high-quality circuits with diversity.Source: §3.1 · p. 2

3.2 Inspiration-Guided Evolution with LLMs

  • IGEL uses the LLM to propose netlists from elite-circuit inspirations, combining it with crossover, mutation, and pruning rather than relying on standalone generation or verification.The LLM receives three rank-sampled elite netlists and their reward scores, then proposes a candidate for SPICE evaluation; netlist-only quality assessment is expected to be unreliable for novel circuits and tasks.Source: §3.2 · pp. 2–3, General Approach · p. 3

General Approach.

  • Removing elite circuits that differ only in parameter values helps preserve topological diversity and reduces premature collapse around a single topology.The authors observed that without this step, the elite set could converge on one topology with minor sizing variations, from which search rarely recovered.Source: Netlist Preprocessing · p. 3

Netlist Preprocessing.

  • Converting PDK-specific transistors into a simplified MOS representation and renaming nets reduces LLM confusion from uncommon names and ambiguous labels.The preprocessing removes explicit bulk connections and clarifies names such as supply and ground nets.Source: Netlist Preprocessing · p. 3

Prompt Templates.

  • Explicit reasoning benefits synthesis, raw-versus-diff prompting matters comparatively little, and alternating reasoning-based raw and diff prompts performs best.Raw prompts generate complete netlists; diff prompts specify changes relative to a base netlist. The alternating scheme’s advantage is attributed to the diversity of its two prompt styles.Source: Prompt Templates · p. 3

4 Analog Discriminant-Function Synthesis

The section defines Iris classification as a non-standard analog synthesis task in which IGEL must infer useful circuit patterns from elite inspirations rather than rely on memorized solutions. It specifies the circuit setup and a multi-corner reward that balances classification, output separation, validity, and circuit size.

4 Analog Discriminant-Function Synthesis

  • The Iris task asks IGEL to infer useful circuit patterns from elite inspirations because it must implement a classifier directly, not reproduce a known circuit class.The discriminant-function synthesis problem is described as non-standard, so the LLM cannot rely on memorized solutions.Source: §4 · p. 3

4.1 Task Setup

  • The classifier uses four normalized Iris features as inputs and three class outputs, with the highest output voltage determining the prediction.Training-set min-max scaling maps feature values to input voltages in the range [0 V, 1.8 V].Source: §4.1 · p. 3
  • The 150 Iris samples are split into 90 training, 30 validation, and 30 test samples, with test data reserved for final evaluation.Training and validation samples are presented in transient simulation; each sample lasts 8 ns, and each split is shuffled three times before concatenation.Source: §4.1 · pp. 3–4
  • Search varies SkyWater transistor topology and device widths and lengths, with a 10 fF load capacitance on each output.Reward also applies a 0.05 penalty per structural validity violation and subtracts 0.0025 × N_transistors to favor smaller circuits.Source: §4.1 · p. 4, §4.2 · p. 4

4.2 Reward and Multi-Corner Evaluation

  • Circuit fitness averages training reward across 17 process, voltage, and temperature corners; averaged validation reward selects the final circuit.The 17 corners comprise nominal tt at 1.8 V and 25°C plus 16 extreme combinations; failed simulations receive a reward of −1.Source: §4.2 · p. 4
  • Reward combines correct predictions with voltage separation, favoring circuits whose correct outputs approach 1.8 V and incorrect outputs approach 0 V.The ideal output is a one-hot voltage vector, and the separation score distinguishes circuits with the same accuracy but different output confidence.Source: §4.2 · p. 4

5 Results

LLM-SPICEMixer augments genetic netlist search with IGEL, an LLM-based proposal operator. Results compare it with evolutionary and machine-learning baselines and assess robustness and design choices.

5 Results

  • Each IGEL step samples three elite circuits as inspirations, obtains one LLM-proposed netlist for SPICE evaluation, and invokes IGEL half as often as each conventional operator.Runs used 131,072 proposal steps; IGEL proposals used reasoning-enabled alternating “raw”/“diff” prompts, balanced decoding, and Qwen3.5 27B.Source: §5 · p. 4

5.1 Comparison with Evolutionary Baselines

  • Median final training reward rose from 0.747 with SPICEMixer to 0.810 with LLM-SPICEMixer (+0.063; p=0.0052), whose best circuit scored 0.855.The median improvement was statistically significant across nine independent runs.Source: Table 1 · p. 5, §5.1 · p. 5
  • Using validation reward to select each run’s best circuit, evaluation on the test split gave LLM-SPICEMixer a median test reward of 0.807, +0.065 over SPICEMixer; three additional strong circuits had diverse topologies while maintaining good performance.Source: §5.1 · p. 5

5.2 Accuracy and Robustness

  • The best circuit achieved 93.4% train, 88.4% validation, and 85.9% test accuracy; under input noise, two leading circuits were competitive with baselines and had higher median accuracy at several larger noise levels.The noise comparison was at the nominal tt setting; overall, synthesized circuits showed degradation comparable to the machine-learning baselines.Source: §5.1 · p. 5, §5.2 · p. 6

5.3 Ablation: Prompting Strategy

  • With Gemma3 12B, reasoning prompts tended to improve median reward, “diff” was slightly more robust than “raw,” and alternating formats with reasoning performed best.The authors suggest edit-based outputs may favor local changes to elite circuits, but report prompt sensitivity and potential gains from better design.Source: §5.3 · p. 6

5.4 Ablation: Decoding Settings

  • For Gemma3 12B, balanced decoding (T=0.7, p_top=0.9) yielded the highest median reward, though differences were limited and other settings were similar.The results suggest moderate stochasticity is a reasonable default, without requiring highly specialized decoding.Source: §5.4 · p. 6

5.5 Ablation: Model Choice and Size

  • Performance depended on model family and size; Qwen3.5 27B had the highest median and maximum reward, improving median reward by +0.05 over Gemma3 27B.Qwen3.5 27B also generated substantially longer responses; the authors note this may indicate useful test-time reasoning but show correlation, not causation.Source: §5.5 · p. 6

5.6 Ablation: Operator Mixture

  • At 18,816 steps, IGEL-only median reward was 0.619 versus 0.689 for the mixture; its best score of 0.656 trailed the mixture’s worst score of 0.665.At 131,072 steps, the mixture’s median reward was 0.763 with the same LLM-call budget; results support using IGEL alongside conventional operators for refinement and search dynamics.Source: §5.6 · p. 6

6 Conclusions and Outlook · A Examples of LLM Prompts and Outputs

LLM-SPICEMixer embeds LLM-generated netlist proposals in a genetic search loop, with SPICE simulation evaluating circuit quality. On Iris discriminant-function synthesis, IGEL improves search performance over SPICEMixer, while the paper identifies generalization, robustness, and proposal strategies as future-work directions.

6 Conclusions and Outlook

  • LLM-SPICEMixer uses IGEL to propose candidate netlists from elite genetic-search solutions, while SPICE simulation remains the source of truth for circuit-quality evaluation.The LLM proposes structured variations rather than generating circuits from scratch.Source: §6 · p. 6
  • IGEL improves search performance over SPICEMixer, reduces premature convergence, and yields compact transistor-level circuits with strong performance on Iris discriminant-function synthesis.The supplied conclusion gives no numerical results for these outcomes.Source: §6 · pp. 6–7
  • Ablations indicate that prompt design and model choice matter in this setting, while decoding configuration has a comparatively smaller effect.Source: §6 · p. 7
  • The training-versus-validation/test reward gap suggests improved reward design or explicit regularization as possible ways to enhance generalization.Source: §6 · p. 7
  • Future synthesis could incorporate noisy or perturbed inputs directly to strengthen robustness, rather than using them only during post-training evaluation.Source: §6 · p. 7
  • Future proposal strategies could explore verbalized sampling [39], self-reflective refinement, or multi-stage proposal-and-repair schemes.Source: §6 · p. 7

A Examples of LLM Prompts and Outputs

B Netlists of the Best Validation Circuits

The section presents the highest-validation-reward LLM-SPICEMixer circuit and additional strong-validation circuits, highlighting its performance across PVT corners and the diversity of discovered solutions.

B Netlists of the Best Validation Circuits

C Output Waveforms Across Shuffles and Corners

Figure 9 compares the input and output waveforms of the two circuits with the best validation performance across operating corners and shuffled test splits.

C Output Waveforms Across Shuffles and Corners

D Robustness to Input-Voltage Perturbations

LLM-SPICEMixer circuits are competitive with ML baselines without input noise and degrade similarly as Gaussian input-voltage perturbations increase. The evaluation reports robustness across noise levels and, for the synthesized circuits, both nominal tt-corner and all-corner performance.

D Robustness to Input-Voltage Perturbations

E Results of Ablation Studies

The ablations compare prompting strategy, decoding settings, model choice, and operator mixture while changing one factor at a time from the default IGEL configuration. They report final best training reward over nine runs; the supplied passages do not include the tables’ numerical results.

E Results of Ablation Studies

  • The ablations vary prompting, decoding, model choice, and operator mixture one factor at a time, reporting final best training reward statistics over nine runs.The reported statistics are average, standard deviation, minimum, median, and maximum; Gemma3 12B was used for Tables 4, 5, and 7 due to GPU resource constraints.Source: E Results of Ablation Studies · p. 15, Table 4 · p. 15, Table 5 · p. 15
  • IGEL-only search is compared with the full operator mixture at 18,816 proposal steps, with the full mixture also reported at 131,072 steps.The IGEL-only budget approximately matches the LLM calls in one full LLM-SPICEMixer run, separating proposal-step budget from LLM-call budget.Source: E Results of Ablation Studies · p. 15, Table 7 · p. 15
  • Table 4 covers prompting strategies, Table 5 decoding settings, and Table 6 model choice for final best reward on the training split.Each table reports values over nine independent runs.Source: Table 4 · p. 15, Table 5 · p. 15, Table 6 · p. 15

F LLM Response Length Statistics

The section reports response-length statistics as additional context for comparing the evaluated LLMs.

F LLM Response Length Statistics

Related papers

Loading 2608.23317v1…