Pith. sign in

REVIEW 4 major objections 5 minor 30 references

AUTOCIRCUIT-RL: Reinforcement Learning-Driven LLM for Automated Circuit Topology Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AUTOCIRCUIT-RL claims that adding PPO-based reinforcement learning with AI reward models on top of instruction-tuned LLMs yields roughly 12% more simulator-valid analog circuit topologies, about 14% higher efficiency, and 38% fewer…

desk verdict Table 1's simulator-verified validity/efficiency gains are real, but the headline success-rate and iterative-adaptation results depend on uncalibrated learned estimators, so those claims are softer than they look. read the letter →

arxiv 2506.03122 v1 pith:LYU6TUUO submitted 2025-06-03 cs.CL

Prashanth Vijayaraghavan , Luyao Shi , Ehsan Degan , Vandana Mukherjee , Xin Zhang This is my paper · ORCID
classification cs.CL
keywords analogcircuittopologysynthesisreinforcementlearningRLAIFpowerconverterdesignlargelanguagemodelsnetlistgenerationrewardmodelingelectronicautomation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a two-phase training recipe—supervised instruction tuning followed by reinforcement learning from AI-generated reward scores—materially improves an LLM's ability to synthesize analog power-converter topologies that satisfy stated design constraints. This matters because analog topology synthesis is normally a slow, simulation-heavy search problem, and the proposed framework performs the refinement only during training, so inference stays fast. On 4- and 5-component circuits, the RL-refined model beats the strongest fine-tuned baselines by about 12% in validity and 14% in efficiency, and with only about 1,000 few-shot examples it extends to 6-10 component circuits while keeping success rates above 60%.

What carries the argument

The load-bearing object is the learned reward function $$r(X,\hat Y) = -1 \text{ if } s_{\text{valid}}<0.6,\; 1 \text{ if } s_{\text{eff}} \text{ or } s_{\text{vout}} \text{ meets constraints},\; s_{\text{eff}} \text{ otherwise},$$ where the scores come from three estimators built on a pretrained transformer text encoder. PPO maximizes this reward while a KL penalty keeps the policy near the instruction-tuned reference model, and the resulting RL-tuned model is then iteratively adapted by sampling its own high-efficiency outputs (efficiency score above 0.7) and retraining on them. That retraining loop is what converts scattered high-quality samples into a stable policy that generates constraint-satisfying circuit topologies directly at inference.

What would settle it

Take the 10,000 circuits selected in iterative adaptation (efficiency score above 0.7) and run them through NGSpice; if their simulator-verified efficiencies are not centered above 0.7, or if the efficiency regressor's ranking correlation with simulator values is near zero on these explored circuits, then the RL phase is optimizing a proxy and the claimed gains would not transfer.

Watch

Extended reading notes

Core claim

The central claim is that optimizing an instruction-tuned LLM with PPO against learned reward models—a circuit-validity classifier, an efficiency regressor, and an output-voltage regressor—pushes generated topologies into the valid, efficient, constraint-satisfying region of design space that random search rarely reaches. The authors' Table 1 reports that AUTOCIRCUIT-RL with Llama-3 reaches 74.48% simulator-validity and 71.65% simulator-efficiency on 4-component circuits, versus 67.32% and 63.87% for the best Gumbel-based baseline, with an overall success rate of 80.69% versus 71.27% and a duplicate generation rate of 1.29 versus 2.19. The reward is sparse and proxy-based: invalid circuits receive -1, circuits meeting efficiency or output-voltage constraints receive 1, and otherwise the efficiency estimate is used as the reward. The paper further claims that iterative adaptation—retraining on self-sampled high-efficiency circuits—contributes substantially, with removing it costing about 8% in efficiency.

Load-bearing premise

The RL stage is only as good as the learned validity and efficiency scores, and the paper never shows that those proxy scores agree with real SPICE simulation on the specific high-efficiency circuits the model explores during iterative adaptation.

Editorial extensions

If this is right

  • At inference the RL-tuned model needs neither reward models nor a SPICE simulator, so valid designs are generated in roughly 2-5 seconds per circuit depending on backbone, far faster than search-based methods that require hundreds of simulation queries.
  • Few-shot fine-tuning on 6-10 component circuits with only about 1,000 examples keeps success rates above 60%, so scaling to larger circuits does not require proportionally larger labeled datasets.
  • Iterative adaptation is doing real work: removing it lowers simulator-verified efficiency by about 8%, meaning self-generated high-efficiency examples are an important part of the gain rather than a negligible add-on.
  • The roughly 38% reduction in duplicate generation rate means fewer samples are wasted when collecting a fixed set of unique topologies, accelerating design-space exploration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same PPO-plus-learned-reward recipe could transfer to other sparse structured-generation tasks such as digital logic synthesis or molecule topology generation, where validity can be learned from simulated examples but exhaustive search is infeasible.
  • A testable consequence of the proxy-reward design is that replacing the efficiency regressor with occasional simulator rollouts in the reward would likely help most on 9-10 component circuits, where the paper's reported success rate declines to about 58.5%.
  • The paper's own error analysis shows most validity failures are small node-labeling mistakes rather than topological errors, so a lightweight rule-based node-correction post-processor could recover many invalid outputs; this is a concrete extension the paper does not test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes AUTOCIRCUIT-RL, a two-phase pipeline for analog circuit topology synthesis: supervised instruction tuning of an LLM on netlists from a random-search dataset, followed by PPO refinement with RoBERTa-based reward estimators for validity, efficiency, and output voltage, plus an iterative-adaptation (IA) stage that selects high-efficiency generations for further RL training. The main empirical claims are a ~12% improvement in validity and ~14% in efficiency over the best baselines on 4- and 5-component power converter circuits, a ~38% reduction in duplicate generation rate, and >60% success in satisfying design constraints with only ~1,000 training examples for 6-10 component circuits. Headline validity and efficiency numbers in Table 1 are partially verified with NGSpice simulation, which is the right kind of evidence; however, the success-rate metric and the IA selection criterion rely on learned estimators trained on the same dataset used for SFT, and the paper does not establish that optimizing these proxies transfers to simulator-verified quality on the model's own generation distribution.

Significance. If the claims hold, the paper makes a useful contribution to LLM-based EDA by showing that reward-driven refinement can improve constraint-satisfying topology generation and that the approach scales to 6-10 components with limited data. Concrete strengths are the use of NGSpice to verify the main validity and efficiency numbers in Table 1, the duplicate-generation-rate analysis, the runtime comparison with search-based methods, and the few-shot generalization experiments. The central weakness is that the evidence for the unique contribution of RL and iterative adaptation is partly circular: the success rate and the IA training-set filter are computed with learned estimators trained on the same random-search data that generated the SFT set. A revision that adds simulator-based success rates and calibration of the estimators on model-generated circuits would considerably strengthen the paper.

major comments (4)
  1. [Section 4.2(d) and Appendix B] The success rate sigma is defined as the percentage of valid topologies that meet design constraints, but the paper does not state whether the constraint check uses NGSpice or the learned estimators. Appendix B says generated topologies are 'analyzed using validity, efficiency, and output voltage estimators' before evaluation, which suggests sigma is estimator-based. Since f_valid, f_eff, and f_vout are trained on the same random-search dataset D used for SFT and reward computation, an estimator-based sigma can be self-confirming and does not demonstrate generalization. Please report sigma computed from NGSpice simulation on the sampled topologies, or provide a calibration plot of estimator scores versus simulator results on model-generated circuits.
  2. [Section 3.2.3 and Section 4.6/Table 1] The iterative-adaptation (IA) training set is selected by nucleus sampling using the learned efficiency estimator with a threshold seff > 0.7, and the reported improvements of AC-RL over AC-RL w/o IA are part of the central claim that RL refinement helps. Although Table 1 reports simulator-based validity and efficiency columns, the 10,000-sample IA selection itself is never checked against NGSpice. If feff is biased on the model's exploration distribution, IA could be selecting circuits that score well under the estimator but not under simulation, and the IA gains could shrink or vanish in other settings. Please report simulator-evaluated statistics for the IA-selected samples and for the final model's outputs, along with a calibration analysis of feff on model-generated circuits.
  3. [Section 4.1 and Table 1] The evaluation omits two baselines that are directly relevant to the claim that RL/IA adds value beyond the training data: (i) random search over the same design space, which is the dataset generator and thus the natural control, and (ii) best-of-k sampling from the SFT or Gumbel-tuned models, which would isolate the effect of reward-guided refinement from the effect of sampling more candidates. Without these comparisons, the ~12% validity and ~14% efficiency gains relative to the Gumbel baseline cannot be cleanly attributed to PPO or iterative adaptation rather than to additional computation or selection effort.
  4. [Table 1 and Section 4.4] All metrics are point estimates from what appears to be a single evaluation run (500 unique topologies per model), with no standard errors, confidence intervals, or significance tests. Given that the headline differences are moderate (e.g., 74.48% vs 67.32% simulator-validity on 4-component circuits), the authors should report variance across random seeds or bootstrap intervals over sampled topologies; otherwise the 12%/14% improvement claims are not statistically grounded.
minor comments (5)
  1. [Section 3.2.1, Circuit Efficiency Estimator] The efficiency estimator is described as a regressor that 'achieves an 83% macro F1 score by categorizing predicted efficiency scores into predefined categories,' which conflates regression and classification; please clarify the estimator type and how the categories are used in the reward.
  2. [Equation (2) and Section 3.2.1] The reward is said to lie in [-1,1], but the clause 'if seff or svout meets constraints' is undefined; please specify the thresholds and how the reward behaves when both efficiency and voltage constraints are present but only one is met.
  3. [Section 4.1 and Table 1] Zero-shot and in-context-learning results are excluded from Table 1 because the outputs are not netlist-like; please provide at least an appendix table or qualitative summary of their failure rates so readers can calibrate task difficulty.
  4. [Throughout] The method name is typeset inconsistently as 'AUTO CIRCUIT -RL', 'AUTOCIRCUIT-RL', and 'AUTOCIRCUIT-RL'; please unify the notation, and likewise use 'success rate' consistently instead of alternating with 'success ratio' in Section 4.3.
  5. [Appendix D] The runtime analysis reports single-point ranges (2.4-5 seconds) without variance or details on sampling configuration; please report mean and standard deviation over repeated generations, including the number of samples.

Circularity Check

2 steps flagged · score 5.0 of 10

Success-rate and iterative-adaptation metrics are gated by the same learned reward estimators; simulator-verified validity/efficiency columns remain independent.

  1. fitted input called prediction [Section 4.2 (Metrics), Appendix B (Inference), Tables 1-2]
    "Success Rate computes the percentage of valid circuit topologies that successfully meet certain design constraints for each category of prompts (as in Section 2): Component constraint (C), Efficiency constraint (C+E), Output V oltage constraint (C+V) and overall success rate (O). [...] These generated topologies are then analyzed using validity, efficiency, and output voltage estimators to identify the designs that are both valid and efficient, meeting the specified design criteria."

    The headline '>60% success' and the σ values in Tables 1-2 are computed for 'valid circuit topologies' with no simulator-based σ column. The only described inference pipeline uses the validity/efficiency/voltage estimators, which are trained on the same synthetic random-search dataset D used for SFT (Section 3.2.1). PPO maximizes exactly these estimators through Eqs. (2)-(3). If σ uses fvalid, feff, and fvout, then a model trained to satisfy those estimators will by construction show high σ regardless of NGSpice truth. The paper reports only aggregate dev-set estimator quality (92% F1, 83% macro F1, MSE 8e-3) and never calibrates the estimators against NGSpice on the model's own generated distribution, so the success-rate claim is not independently grounded.

  2. fitted input called prediction [Section 3.2.3 (Iterative Adaptation), Eqs. (2)-(3)]
    "Utilizing the nucleus sampling approach, we specifically target circuit topologies with an efficiency score ( seff) exceeding 0.7. This criterion ensures that the selected circuits not only satisfy validity constraints but also exhibit high operational efficiency. [...] we iteratively refine the model using Equation 3, incorporating insights from high-quality topologies into the RL-refined model from the previous iteration."

    The IA phase selects 'high-quality' samples using the learned efficiency regressor feff (seff > 0.7) and then fine-tunes with the reward r(X, Y-hat) defined through the same seff in Eq. (2). Thus the 10,000-sample IA dataset is 'high-quality' only relative to feff, not to NGSpice. The reported ~8% efficiency gain from IA (Section 4.6) is therefore partly self-confirming if measured with feff; the simulator columns in Table 1 do show an IA gain for the main comparison, but no simulator calibration of feff on the IA-selected distribution is provided, so the direction and magnitude of the learned-proxy bias are unknown.

full rationale

The central validity and efficiency comparisons are independently grounded: Table 1 reports NGSpice simulator columns E(fSvalid) and E(fSeff), and AC-RL (Llama-3) outperforms the best Gumbel baseline on 4-component circuits (74.48% vs 67.32% validity; 71.65% vs 63.87% efficiency). So the main '~12% more valid circuits' and '~14% efficiency' claims are not pure fit artifacts, and there is no load-bearing self-citation chain. However, two quantities used to support the abstract and the RL ablation are gated by the same learned estimators trained on the same dataset D: the success rate σ (abstract's 'over 60% success') and the iterative-adaptation sample selection. The paper never reports a simulator-computed success rate or a calibration of the estimators on model-generated circuits, so those specific claims are partially self-confirming. The score of 5 reflects this partial circularity while acknowledging the independent simulator-verified columns in Table 1.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The framework rests on the random-search-generated dataset and on AI reward models trained on that same data. The main free choices are the data-group sampling weights, the validity threshold of 0.6, the iterative-adaptation efficiency threshold of 0.7, and the unstated KL penalty eta. No new physical entities are introduced.

free parameters (5)
  • data group sampling weights = Group1: 0.1, Group2: 0.25, Group3: 0.25, Group4: 0.4
    Chosen by hand to enrich training batches with high-efficiency circuits (Section 2, Appendix B); directly shapes the SFT and RL data distribution.
  • validity threshold = 0.6
    Used in the reward function (Eq. 2, reward = -1 if svalid < 0.6) and in the metric definition (Section 4.2: a circuit is valid if its classifier validity score exceeds 0.6). This hand-set threshold determines what counts as valid.
  • iterative adaptation efficiency threshold = 0.7
    Circuits with estimated efficiency exceeding 0.7 are selected for iterative refinement (Section 3.2.3); the value 0.7 is chosen without stated justification.
  • KL penalty coefficient eta = not reported
    In Eq. (3), eta controls the KL penalty in PPO; the paper never states its value, so the RL results cannot be reproduced as-is.
  • evaluation duty cycles = 0.1, 0.3, 0.5, 0.7, 0.9
    Each netlist is simulated at these five duty cycles (Section 2); the selection rule for the reported performance value is not stated.
assumptions (6)
  • domain assumption NGSpice simulation accurately approximates the real electrical behavior of the generated circuits.
    All validity, efficiency, and output-voltage ground truth comes from NGSpice with fixed device models; if the simulation is inaccurate, the true quality of circuits is unknown. Section 2.
  • domain assumption The simplified device model (FETs with fixed gate/body wiring, fixed capacitance 10uF, inductance 10uH, 200kHz, Vin=2V) captures the design problem of interest.
    The entire dataset and all evaluations use these fixed parameters; results may not generalize to other component values or device families. Section 2.
  • domain assumption The incident encoding (Fatemi et al., 2024) is an adequate representation for LLM generation of netlists.
    The paper adopts incident encoding for all models without comparing to other encodings in this paper. Section 2.
  • ad hoc to paper The reward estimators (RoBERTa validity classifier, efficiency regressor, voltage regressor) are accurate enough proxies for NGSpice.
    The entire RL signal is generated by these estimators trained on the same dataset; only dev-set F1/MSE are reported, not estimator-simulator agreement on selected samples. Section 3.2.1.
  • standard math Standard PPO and RLAIF training procedures are assumed correct and stable as implemented.
    The paper follows Ouyang et al. (2022) and Schulman et al. (2017) without modification. Section 3.2.2.
  • ad hoc to paper The four-group categorization of training data with hand-picked boundaries is a valid stratification for learning.
    Groups 1-4 are defined by efficiency and voltage ranges with arbitrary thresholds (e.g., efficiency less than 0.05, 0.05 to 0.7); the weighted sampling may bias the model toward Group 4. Section 2, Appendix B.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AUTOCIRCUIT-RL: Reinforcement Learning-Driven LLM for Automated Circuit Topology Generation." pith.science (2026). https://pith.science/paper/LYU6TUUO

@misc{pith2026250603122,
  author       = {Pith},
  title        = {Pith review of: AUTOCIRCUIT-RL: Reinforcement Learning-Driven LLM for Automated Circuit Topology Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LYU6TUUO}},
  note         = {Machine review of arXiv:2506.03122}
}
read the original abstract

Analog circuit topology synthesis is integral to Electronic Design Automation (EDA), enabling the automated creation of circuit structures tailored to specific design requirements. However, the vast design search space and strict constraint adherence make efficient synthesis challenging. Leveraging the versatility of Large Language Models (LLMs), we propose AUTOCIRCUIT-RL,a novel reinforcement learning (RL)-based framework for automated analog circuit synthesis. The framework operates in two phases: instruction tuning, where an LLM learns to generate circuit topologies from structured prompts encoding design constraints, and RL refinement, which further improves the instruction-tuned model using reward models that evaluate validity, efficiency, and output voltage. The refined model is then used directly to generate topologies that satisfy the design constraints. Empirical results show that AUTOCIRCUIT-RL generates ~12% more valid circuits and improves efficiency by ~14% compared to the best baselines, while reducing duplicate generation rates by ~38%. It achieves over 60% success in synthesizing valid circuits with limited training data, demonstrating strong generalization. These findings highlight the framework's effectiveness in scaling to complex circuits while maintaining efficiency and constraint adherence, marking a significant advancement in AI-driven circuit design.

Figures

Figures reproduced from arXiv: 2506.03122 by the authors.

Figure 1
Figure 1. Given a prompt with design constraints, the goal is to gen￾erate a circuit topology and duty cycle satisfying those constraints. Component constraints (blue) are mandatory, while efficiency (or￾ange) and voltage (purple) constraints are optional. In the output, blue circles denote components (“F” for FET, “I” for inductor), and yellow circles denote connection nodes. 1. Introduction AI and machine learning have been… view at source ↗
Figure 2
Figure 2. ((a) Statistics of our Circuit dataset. b) Vout-Efficiency graph of 4-component circuits in the training data. (c) Illustration of the AUTOCIRCUIT-RL framework. Dashed lines depict probability flow, while the dotted line represents iterative adaptation post RLAIF tuning. A prompt guides topology generation, then rewards are calculated with KL-divergence to a reference model to preserve the original distribution. PPO… view at source ↗
Figure 3
Figure 3. RL convergence curves for AUTOCIRCUIT-RL with Llama-3 over ∼25,000 training steps. Top: Efficiency for 4- and 5-component circuits; Bottom: Success ratio of valid and constraint-satisfying topologies. ∼ 2.7 seconds per design generation once trained (refer Section D for runtime analysis). In contrast, traditional search-based algorithms (Fan et al., 2021) are computation￾ally expensive and time-consuming, often taki… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Few-shot fine-tuning results for 6-10 components, with models tuned using k-examples of varying circuit topologies [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Plot of SuccessRate@m of AC-RL for m ∈ {1, 3, 5}. Efficiency (E) & Output Voltage (V) Adherence Meeting efficiency and expected output voltage constraints is chal￾lenging in comparison to the component pool constraint, par￾ticularly when expected values are specified i…
Figure 6
Figure 6. Figure 6: Vout-Efficiency graph of 5-component circuits in the training data. • Iterative Adaptation: The iterative adaptation, explained in Section 3.2.3, is performed over 3–5 iterations depending on the base model used. Each iteration involves 2–4 epochs of RL refinement and …
Figure 7
Figure 7. Figure 7: Runtime Complexity of AUTOCIRCUIT-RL Using MPT-7B vs. LLaMA-3 8B. The runtime per circuit increases with the number of components but remains significantly faster than traditional search-based methods. MPT-7B offers lower latency due to its smaller model size, while LL…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 15 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Constitutional ai: Harmlessness from ai feedback

    Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., Chen, A., Goldie, A., Mirhoseini, A., McKinnon, C., et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022

    arXiv 2022

  3. [3]

    Estimating or propagating gradients through stochastic neurons for conditional computation

    Bengio, Y., L \'e onard, N., and Courville, A. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013

    arXiv 2013

  4. [4]

    Auto-spice: Leveraging llms for dataset creation via automated spice netlist extraction from analog circuit diagrams

    Bhandari, J., Bhat, V., He, Y., Garg, S., Rahmani, H., and Karri, R. Auto-spice: Leveraging llms for dataset creation via automated spice netlist extraction from analog circuit diagrams. arXiv preprint arXiv:2411.14299, 2024

    arXiv 2024

  5. [5]

    GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow , March 2021

    Black, S., Leo, G., Wang, P., Leahy, C., and Biderman, S. GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow , March 2021. URL https://doi.org/10.5281/zenodo.5297715

    doi:10.5281/zenodo.5297715 2021

  6. [6]

    Lamagic: Language-model-based topology generation for analog integrated circuits

    Chang, C.-C., Shen, Y., Fan, S., Li, J., Zhang, S., Cao, N., Chen, Y., and Zhang, X. Lamagic: Language-model-based topology generation for analog integrated circuits. In the 41st International Conference on Machine Learning (ICML), pp.\ 1--8, 2024

    work page 2024

  7. [7]

    Artisan: Automated operational amplifier design via domain-specific large language model

    Chen, Z., Huang, J., Liu, Y., Yang, F., Shang, L., Zhou, D., and Zeng, X. Artisan: Automated operational amplifier design via domain-specific large language model. In Proceedings of the 61st ACM/IEEE Design Automation Conference, pp.\ 1--6, 2024

    work page 2024

  8. [8]

    Cktgnn: Circuit graph neural network for electronic design automation

    Dong, Z., Cao, W., Zhang, M., Tao, D., Chen, Y., and Zhang, X. Cktgnn: Circuit graph neural network for electronic design automation. In International Conference on Learning Representations (ICLR), pp.\ 1--20, 2023

    work page 2023

Show all 30 references
  1. [9]

    From specification to topology: Automatic power converter design via reinforcement learning

    Fan, S., Cao, N., Zhang, S., Li, J., Guo, X., and Zhang, X. From specification to topology: Automatic power converter design via reinforcement learning. In IEEE/ACM International Conference On Computer Aided Design (ICCAD), pp.\ 1--9, 2021

    2021

  2. [10]

    Talk like a graph: Encoding graphs for large language models

    Fatemi, B., Halcrow, J., and Perozzi, B. Talk like a graph: Encoding graphs for large language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=IuXR1CCrSi

    2024

  3. [11]

    The llama 3 herd of models

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  4. [12]

    Bagnet: Berkeley analog generator with layout optimizer boosted with deep neural networks

    Hakhamaneshi, K., Werblun, N., Abbeel, P., and Stojanovi \'c , V. Bagnet: Berkeley analog generator with layout optimizer boosted with deep neural networks. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pp.\ 1--8, 2019

    2019

  5. [13]

    Categorical reparameterization with gumbel-softmax

    Jang, E., Gu, S., and Poole, B. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016

    2016 arXiv

  6. [14]

    Z., and Luo, P

    Lai, Y., Lee, S., Chen, G., Poddar, S., Hu, M., Pan, D. Z., and Luo, P. Analogcoder: Analog circuit design via training-free code generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 379--387, 2025

    2025

  7. [15]

    Rlaif: Scaling reinforcement learning from human feedback with ai feedback

    Lee, H., Phatale, S., Mansoor, H., Lu, K., Mesnard, T., Bishop, C., Carbune, V., and Rastogi, A. Rlaif: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023

    2023 arXiv

  8. [16]

    The power of scale for parameter-efficient prompt tuning

    Lester, B., Al-Rfou, R., and Constant, N. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021

    2021 arXiv

  9. [17]

    and Hutter, F

    Loshchilov, I. and Hutter, F. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  10. [18]

    Trustworthy Genetic Programming - Based Synthesis of Analog Circuit Topologies Using Hierarchical Domain - Specific Building Blocks

    McConaghy, T., Palmers, P., Steyaert, M., and Gielen, G. Trustworthy Genetic Programming - Based Synthesis of Analog Circuit Topologies Using Hierarchical Domain - Specific Building Blocks . IEEE Transactions on Evolutionary Computation, 15 0 (4): 0 557--570, 2011. ISSN 1941-0026

    2011

  11. [19]

    and Vogt, H

    Nenzi, P. and Vogt, H. Ngspice users manual version 23. Experiments/ ngspice23-manual. pdf, 2011

    2011

  12. [20]

    Training language models to follow instructions with human feedback

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 27730--27744, 2022

    2022

  13. [21]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Reimers, N. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019

    1908 arXiv

  14. [22]

    Proximal policy optimization algorithms

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  15. [23]

    Atelier: An automated analog circuit design framework via multiple large language model-based agents

    Shen, J., Chen, Z., Zhuang, J., Huang, J., Yang, F., Shang, L., Bi, Z., Yan, C., Zhou, D., and Zeng, X. Atelier: An automated analog circuit design framework via multiple large language model-based agents. Authorea Preprints, 2024

    2024

  16. [24]

    and Komodakis, N

    Simonovsky, M. and Komodakis, N. Graphvae: Towards generation of small graphs using variational autoencoders. In Artificial Neural Networks and Machine Learning--ICANN 2018: 27th International Conference on Artificial Neural Networks, Rhodes, Greece, October 4-7, 2018, Proceed...

    2018

  17. [25]

    Team, M. N. et al. Introducing mpt-7b: A new standard for open-source, commercially usable llms, 2023

    2023

  18. [26]

    Circuitsynth: Leveraging large language models for circuit topology synthesis

    Vijayaraghavan, P., Shi, L., Degan, E., and Zhang, X. Circuitsynth: Leveraging large language models for circuit topology synthesis. In 2024 IEEE LLM Aided Design Workshop (LAD), pp.\ 1--6. IEEE, 2024

    2024

  19. [27]

    Gcn-rl circuit designer: Transferable transistor sizing with graph neural networks and reinforcement learning

    Wang, H., Wang, K., Yang, J., Shen, L., Sun, N., Lee, H., and Han, S. Gcn-rl circuit designer: Transferable transistor sizing with graph neural networks and reinforcement learning. In ACM/IEEE Design Automation Conference (DAC), pp.\ 1--6, 2020

    2020

  20. [28]

    Analogxpert: Automating analog topology synthesis by incorporating circuit design expertise into large language models

    Zhang, H., Sun, S., Lin, Y., Wang, R., and Bian, J. Analogxpert: Automating analog topology synthesis by incorporating circuit design expertise into large language models. arXiv preprint arXiv:2412.19824, 2024

    2024 arXiv

  21. [29]

    D-vae: A variational autoencoder for directed acyclic graphs

    Zhang, M., Jiang, S., Cui, Z., Garnett, R., and Chen, Y. D-vae: A variational autoencoder for directed acyclic graphs. Advances in neural information processing systems, 32, 2019

    2019

  22. [30]

    and Zhang, L

    Zhao, Z. and Zhang, L. An Automated Topology Synthesis Framework for Analog Integrated Circuits . IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 39 0 (12): 0 4325--4337, 2020. ISSN 1937-4151

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.