acceptodds
Under review as a conference paper at ICLR 2027

RGB-ACD: RGB-Inspired Topology Generation for LLM-Guided Analog Circuit Design

Abstract

Automating analog circuit design requires searching discrete topologies and continuous device parameters under coupled performance constraints. Translating natural-language requirements into circuit structures that satisfy these constraints remains challenging. We introduce RGB-inspired Analog Circuit Design (RGB-ACD), a large language model (LLM)-guided agent for topology generation and device sizing. RGB-ACD represents circuit topologies as three-channel RGB images and uses this encoding to learn a mapping from text to circuit connectivity. We construct a dataset of 35,200 text-topology pairs and train a generator that predicts electrical connections while preserving fixed device structure. To meet performance targets, the agent couples an outer topology-refinement loop with an inner parameter-optimization loop. The outer LLM uses simulation residuals to revise topology-generation instructions while retaining the original targets. For each candidate topology, the inner loop uses finite-difference gradient estimates to update sensitive parameter groups under process constraints. RGB-ACD outperforms all evaluated baselines in success rate across all three difficulty levels on OCB:Ckt-Bench-101 and AMSBench-Design. On hard tasks, success rates reach 66% and 51%, exceeding the strongest baselines by 42 and 30 percentage points, respectively.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Analog Circuit Design, LLM Agents, Topology Generation, RGB Representation, Simulation-Guided Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_xabtOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RGB-ACD combines a three-channel circuit-connectivity representation, text-conditioned topology generation, LLM-directed refinement, and finite-difference device sizing. Trained on 35,200 text-topology pairs, it is evaluated on 2,200 tasks across two benchmarks, reporting hard-task success rates of 66% and 51%, exceeding the strongest evaluated baselines by 42 and 30 percentage points.

Critical strengths
  1. Separating fixed device structure from predicted electrical connections is sensible. Preserving original specifications and requiring final simulation verification make the design objective concrete.

  2. Large reported improvements across two benchmarks and all difficulty levels suggest practical potential. The representation ablation provides supporting evidence for separating connection types.

Critical weaknesses
  1. Generalization and task scope are unclear. Topology-level separation among training, retrieval, and test data is not documented, and test-time device-inventory selection is unspecified despite being required by the generator.

  2. The comparisons do not isolate feedback effectiveness or computational efficiency. Removing feedback also reduces search to one attempt, versus up to fifteen for the full method. Baseline simulation budgets are unspecified, and reported attempt counts do not distinguish inner updates from outer attempts.

Questions for authors
  1. Are device inventories inferred from requests or supplied from reference circuits?

  2. How is training/retrieval/test separation enforced, including storage of previously solved test tasks?

  3. Do feedback gains persist against a fifteen-attempt baseline without residual feedback, using identical simulation budgets?

3 comments

Thank you for recognizing the separation of fixed structure from predicted connectivity and the simulation-verification pipeline. We address test-time device configuration, data separation, and the feedback comparison below.

R4.1. Test-time generation uses a shared fixed template with task-independent device availability

We use the fixed template in Supplementary File 1, referenced in Section 2.2.2. It provides 1,024 device-terminal/port indices and an initially empty connectivity matrix. Every test task and generation attempt activates the same predefined set of template positions. The active-node indicator a remains fixed, and the available device/port positions are independent of the task.

Within this shared template, the generator preserves Grid and predicts Routing and Bridging to determine each candidate circuit’s connectivity.

All methods in Table 2 receive the same device/template information and share the retrieval library and retrieval permissions. They use the same inner sizing optimizer, with K=15, T=10, and at most 400 Ngspice calls per candidate. Each method stops at the first independently verified feasible design; unsuccessful tasks exhaust 15 topology attempts.

R4.2. Splitting precedes augmentation, and retrieval-library updates are disabled during evaluation

Data are partitioned by original circuit ID before augmentation. Samples produced through device substitutions, specification variations, and text paraphrases inherit their source circuit’s split assignment. All augmented samples derived from the same original circuit remain in the same split. Training circuits come from AnalogGenie (Appendix B), and the evaluation benchmarks are derived from OCB and AMSBench (Appendix D).

The retrieval library is built from a separately prepared circuit collection and separated from test data before evaluation. The successful-design storage operation in Figure 3 and Algorithm 2 applies to day-to-day use and is disabled during evaluation. Designs produced by test tasks therefore do not enter the retrieval library for subsequent test tasks.

R4.3. Feedback achieves higher success rates at the same topology-attempt limit

On the complete Hard splits of OCB:Ckt-Bench-101 and AMSBench-Design, RGB-ACD achieves success rates of 66% / 51%, compared with 57.5% / 43% for up to 15 independent restarts without feedback. Both methods allow at most 15 topology attempts, with the feedback version achieving gains of 8.5 and 8 percentage points, respectively.

MethodMaximum topology attemptsOCB:Ckt-Bench-101 Hard SRAMSBench-Design Hard SR
RGB-ACD with feedback1566%51%
Independent restarts without feedback1557.5%43%
Original no-feedback ablation in Table 3 (single topology)143%28%

My concerns have been addressed, and I would like to maintain my positive score.

Thank you for your follow-up and continued support. If you have any further questions, we would be happy to address them.

Reviewer_gSaKOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes RGB-ACD, an LLM-guided agent for analog circuit topology generation and device sizing. Circuit connectivity is encoded as three aligned binary relation maps on a fixed 1024×1024 node index: Grid (device-to-terminal incidence, fixed), Routing (terminal-to-terminal connections), and Bridging (terminal-to-port connections). Also, it proposes a black-box optimization engine to search for device sizing parameter. Then, these tools are provided as an MCP for agent to search for circuit and do optimization.

Critical strengths

RGB model is a novel idea to encode connections in the circuit into 2D grid. Agent system is comprehensive to use the proposed tools to help circuit optimization. The paper evaluates on two well-known benchmarks and run the simulation results rigorously, including direct-current (DC), alternating-current (AC), transient, and noise analyses

Critical weaknesses

The main weakness is the motivation of using CNN is not clear. Please provide comparison with other GNN and transformer works such as LaMAGIC2 and AnalogGene. The paper claims that the other previous works do not do sizing and thus the paper doesn't include the comparison. However, in the paper, the framework do topology generation first than sizing. Therefore, it is important to provide comparison that isolates the performance of topology generation and sizing.

Questions for authors
  1. Table 2 does not say whether ChatGPT-5.5, DeepSeek-V4-Pro, and Qwen3-32B (RAG) had access to the Ngspice sizing tool, the retrieval library, or the same K/T and simulation budgets, so the gap may come from the tooling rather than the RGB representation. Since Algorithm 1 accepts any topology τ, please run AnalogGenie and LaMAGIC as the topology module inside the same agent (same LLM, retrieval, sizing loop, budgets) so that we can evaluate your proposed RGB model. Note that LaMAGIC has a follow-up work LaMAGIC2, so you can compare with it.
  2. The model applies 3×3 convolutions over a 1024^2 canvas that is ~98% empty slots. It could approximately only have a ~10⁻⁴ positive rate, and the loss includes those empty entries. Can you compare your RGB model along with node-level GNN or transformer model to show the effectiveness of your proposed model?
  3. You can have additional experiment to show the benefit of your RGB model. (1) for each benchmark task, take the ground-truth topology the task was built from, skip generation, and run only the sizing loop -> we can know the upper bound of topology generation and your sizing algorithm effectiveness. (2) Use the same agent but replace the RGB generator with the top circuit returned by retrieval (component A), run sizing and the outer loop as usual -> this shows that whether the generator can propose new circuit rather than just copying the top-performance circuit.
  4. Please report the overlap between the 35,200 AnalogGenie-derived training circuits and the 2,200 test circuits. We would like to validate the generalization ability of RGB mode. Also, you can report the token cost compared to other methods.
8 comments
AuthorsOpenReview ↗

Response to Reviewer gSaK — Part 1/2

Thank you for recognizing the relation-based representation, integrated design system, and simulation-based evaluation. We address topology-module comparisons, architecture choice, retrieval and reference-topology experiments, and data separation below.

R3.1. Shared sizing tools and topology-module comparisons

All methods in Table 2 use the same sizing optimizer, input information, and budget limits. AnalogSeeker and all LLM baselines use the inner sizing optimizer in Algorithm 1 and share the device/template information, retrieval library, and retrieval permissions. The budgets are K=15, T=10, and at most 400 Ngspice calls per candidate. All methods stop at the first independently verified feasible design; unsuccessful tasks exhaust 15 topology attempts.

AnalogGenie addresses circuit-topology discovery; its original paper does not demonstrate the natural-language-instruction-driven, open-ended design workflow considered here.

The additional topology-module comparison evaluates the LaMAGIC2 replacement on the complete Hard splits of both benchmarks. The RGB generator achieves success rates 12 and 9 percentage points higher on OCB:Ckt-Bench-101 Hard and AMSBench-Design Hard, respectively.

Topology moduleOCB:Ckt-Bench-101 Hard SRAMSBench-Design Hard SR
RGB generator (full method)66%51%
LaMAGIC254%42%

R3.2. Relation representation on fixed indices and architecture comparisons

The RGB representation models fixed Grid relations separately from the Routing and Bridging connections to be predicted. Matrix coordinates correspond to device-terminal/port indices in Supplementary File 1, with consistent node ordering across samples. These coordinates represent connectivity relations rather than physical layout positions (Appendix A.2). The CNN predicts connections on this relation matrix through shared convolutional parameters.

We evaluate a relational GNN, a Transformer, and a variant jointly removing self-loop and soft-degree regularization on the complete Hard splits.

ConfigurationOCB:Ckt-Bench-101 Hard SRAMSBench-Design Hard SR
RGB-ACD (CNN)66%51%
Relational GNN35%37%
Transformer65%50.5%
Joint removal of self-loop and soft-degree regularization53%45%

The CNN achieves the highest success rates in the reported architecture comparisons, with the Transformer reaching a similar level. The full model outperforms the variant jointly removing both regularizers by 13 and 6 percentage points on OCB:Ckt-Bench-101 Hard and AMSBench-Design Hard, respectively.

AuthorsOpenReview ↗

Response to Reviewer gSaK — Part 2/2

R3.3. Comparisons of generated, retrieved, and reference topologies

The full method exceeds the Top-1 retrieval variant with multiple sizing initializations by 28.5 and 26 percentage points, respectively. We apply sizing to the top-ranked retrieved circuit and to the benchmark reference topology to evaluate direct reuse of retrieved circuits and sizing performance with a supplied reference topology. All results below cover the complete Hard splits of both benchmarks.

Topology source and sizing settingOCB:Ckt-Bench-101 Hard SRAMSBench-Design Hard SR
RGB-generated topology (full method)66%51%
Top-1 retrieved circuit; single sizing run30%19%
Top-1 retrieved circuit; multiple sizing initializations37.5%25%
Benchmark reference topology; single sizing run76%64%
Benchmark reference topology; multiple sizing initializations84%73%

Generated topologies achieve higher task success rates than direct retrieval in these evaluations. The reference-topology experiment provides a diagnostic of the sizing stage: a single run achieves 76% / 64%, and multiple initializations achieve 84% / 73%.

R3.4. Data separation

Training data come from AnalogGenie and contain 35,200 text–topology pairs (Appendix B); the evaluation benchmarks are derived from OCB and AMSBench (Appendix D). Data are partitioned by original circuit ID before augmentation, and all augmented samples from the same original circuit remain in the same split. There are 3,350 unique source circuits before augmentation. Our check found no structural overlap between circuits with different original IDs.

The retrieval library comes from a separately prepared circuit collection and is separated from evaluation data in advance. Successful-design insertion applies to day-to-day use and is disabled during evaluation, so later test tasks cannot retrieve designs generated by earlier test tasks.

Dear Reviewer gSaK,

We wanted to follow up on the additional comparisons in Parts 1 and 2 of our response. We would be glad to hear your thoughts on the results and clarify any remaining concerns. If these results address your main concerns, we hope you will reconsider your assessment and rating.

Thank you for your time.

Reviewer_BvUWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work focuses on automating analog circuit design. It builds a LLM-guided agent for topology generation and device sizing, called RGB-ACD. It represents circuit topologies as RGB images and use it to learn a mapping from text to circuit connectivity. It also couples an outer topology-refinement loop with an inner parameter-optimization loop.

Critical strengths
  • This paper focuses an practical and valuable problem, which has not been well-solved by LLM.
  • Using image to represent circuit topologies is interesting, as well as introducing CLIP for RGB topology generation.
  • The overall design of agent framework is reasonable.
Critical weaknesses
  • How the dataset (the 35k text-topology pairs) is collected or synthesized? And, could you add some statistics to demonstrate its diversity and distribution?
  • In the ablation study, what does "single-layer topology encoding" mean? Could you briefly introduce how it is implemented? And, is it possible to compare the proposed image representation with the existing ones illustrated in Figure 2 with experimental results?
  • Could RGB-ACD framework also work well using open-source LLMs (e.g., Kimi, GLM) and lightweight LLMs (e.g., Qwen 30B series)?
Questions for authors

See weaknesses

3 comments

Thank you for recognizing the practical value of the problem, the RGB representation, and the overall agent design. We address dataset diversity, representation and architecture comparisons, and outer-loop LLM replacement below.

R2.1. Dataset construction and diversity

Our dataset contains 35,200 text–topology pairs. As described in Appendix B, we convert AnalogGenie circuit structures into relation maps and construct descriptions using circuit images, Cadence schematics, adjacency matrices, and specification tables. Descriptions are manually checked for consistency with circuit structures and numerical specifications. Augmentation includes functionally equivalent device substitutions, specification variations within a topology family, and textual paraphrases.

The AnalogGenie source collection contains 3,350 unique topologies across 11 functional categories. These cover amplification, regulation, reference generation, comparison, oscillation, frequency conversion, power conversion, and sampling. The largest categories are operational amplifiers (836; 24.96%), switched-capacitor samplers (723; 21.58%), and bandgap references (667; 19.91%). Circuit sizes also vary: 56.45% contain more than 20 devices, and 17.07% contain more than 30.

Data are partitioned by original circuit ID before augmentation, so all variants of the same source circuit remain in the same split. Evaluation data are separated from training and retrieval data. The retrieval library is prepared separately and remains fixed during evaluation; successful test designs are not inserted.

R2.2. Contribution of the representation and comparisons with alternative architectures

Table 3 reports gains of 35 and 25 percentage points for the multi-layer representation on the two Hard splits. RGB-ACD achieves success rates of 66% and 51%, compared with 31% and 26% for the single-layer variant.

The single-layer variant uses the same template and node indexing, while combining connection types into one layer. Fixed Grid constraints, the active-node indicator a, the generator backbone, training data and settings, outer-loop feedback, retrieval, inner sizing optimizer, and evaluation budgets remain unchanged. Both representations use the same matrix-to-netlist decoding rules.

The additional comparison evaluates a relational GNN as a predictor-architecture replacement:

ConfigurationComparisonOCB:Ckt-Bench-101 Hard SRAMSBench-Design Hard SR
RGB-ACDFull method66%51%
Single-layer representationConnection-type encoding31%26%
Relational GNNPredictor-architecture replacement35%37%

R2.3. Replacing the outer-loop LLM

The controller comparison replaces the outer-loop LLM within RGB-ACD. The topology generator, retrieval library, inner sizing optimizer, and evaluation budgets are held fixed to isolate controller dependence.

Outer-loop LLMOCB:Ckt-Bench-101 Hard SRAMSBench-Design Hard SR
GPT-5.5, as used in the paper66%51%
Qwen3-32B0%0%
Kimi K3.1 Pro63%52%

Dear Reviewer BvUW,

We would be happy to clarify any remaining questions. If you feel that our response has addressed your main concerns, we would be grateful if you could reconsider your assessment and rating.

Thank you for your consideration.

We hope the additional experiments help clarify the value of our approach. The single-layer ablation supports separating circuit relations, and the Kimi results show that the system also works with another LLM controller. If these findings address your main concerns, we would appreciate your reconsideration of the recommendation. We would be happy to discuss any remaining questions.

Reviewer_ej1uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RGB-ACD addresses natural-language-guided analog circuit design by combining conditional connectivity generation with simulation-based device sizing. It represents circuits using three relation maps: Routing for inter-device electrical connections, Grid for fixed device-to-terminal incidence, and Bridging for terminal-to-port connections. A frozen CLIP text encoder conditions a FiLM-based generator that predicts Routing and Bridging while preserving Grid. An outer LLM loop revises design instructions using simulation diagnostics, while an inner finite-difference optimizer adjusts device parameters under fixed specifications and PDK bounds. The paper constructs 35,200 text-topology pairs and evaluates the system on OCB:Ckt-Bench-101 and an AMSBench-derived design benchmark. It reports hard-task success rates of 66% and 51%, respectively, exceeding the strongest evaluated baselines by 42 and 30 percentage points.

Critical strengths
  1. The separation of fixed device structure from predicted electrical connectivity is a well-motivated modeling choice (Sections 2.2-2.3). Preserving Grid avoids predicting known device incidence, while symmetrization enforces undirected connectivity. This provides a clear structural inductive bias for circuit generation.

  2. The reported end-to-end improvements are substantial across both benchmarks and all difficulty levels (Table 2). The system integrates topology generation and sizing, and acceptance requires final simulation to verify all original performance constraints. This makes the evaluation more practically relevant than topology validity alone, subject to the protocol concerns below.

Critical weaknesses
  1. The ablations do not adequately isolate the contributions of feedback and the representation (Section 3.5, Table 3). The full method permits up to 15 outer attempts, whereas the no-feedback variant reports ANoA = 1.0 in every setting and substantially lower runtime. Thus, the success-rate improvement conflates feedback-guided refinement with additional search opportunities and computation. A no-feedback restart baseline with matched topology attempts and simulator-call budgets is needed to establish the value of feedback itself. Similarly, the single-layer ablation is insufficiently specified to determine whether fixed-structure constraints, model capacity, and training settings are controlled. These limitations weaken the attribution of the gains to the proposed components.

  2. Important evaluation details remain unclear in the manuscript, limiting confidence in generalization and comparison fairness. The generator receives a device template and active-node set, but inventory selection lies outside the generator (Section 2.3.1 and Appendix E); the test-time source of these inputs and equivalent access for baselines need clarification. Appendix B describes topology-preserving augmentation, but does not establish topology- or family-disjoint separation between training, retrieval, and evaluation data. The protocol for storing successful test designs in the retrieval library also needs clarification. Finally, baseline simulation budgets and runtime aggregation are not sufficiently specified, and Section 3.3 explicitly notes that ANoA does not distinguish outer attempts from inner parameter updates. These are unresolved protocol questions rather than evidence of demonstrated leakage, but they materially affect the interpretation of the reported improvements.

Questions for authors
  1. Can you compare feedback-guided refinement against no-feedback restarts under the same number of topology attempts and the same total simulator-call budget? This would help distinguish the benefit of diagnostic feedback from additional search.

  2. How are the device template and active-node set selected at test time? Are they derived from reference circuits, and do all baselines receive equivalent information?

  3. How are training examples, retrieval references, and evaluation tasks separated at the topology or topology-family level, including augmented variants? Is the retrieval library reset between evaluation tasks, or can earlier test solutions affect later tasks?

  4. Please specify the simulator-call budgets and tool access for each baseline, how runtime includes successful and failed tasks, and whether the single-layer ablation preserves fixed-structure constraints, model capacity, and training settings. These clarifications and controls could materially change the assessment.

5 comments

Thank you for recognizing the representation design and simulation-verification pipeline. We address your questions about feedback, test-time inputs, data separation, and controlled evaluation below.

R1.1. Feedback achieves higher success rates at the same topology-attempt limit

On the complete Hard splits of OCB:Ckt-Bench-101 and AMSBench-Design, RGB-ACD achieves success rates of 66% and 51%, compared with 57.5% and 43% for up to 15 independent restarts without feedback. Both methods allow at most 15 topology attempts, with the feedback version achieving gains of 8.5 and 8 percentage points, respectively.

MethodMaximum topology attemptsOCB:Ckt-Bench-101 Hard SRAMSBench-Design Hard SR
RGB-ACD with feedback1566%51%
Independent restarts without feedback1557.5%43%
Original no-feedback ablation in Table 3143%28%

R1.2. Test-time generation uses a shared fixed template

We use the fixed template provided in Supplementary File 1, referenced in Section 2.2.2. It defines 1,024 device-terminal/port indices and an initially empty connectivity matrix. Every test task and generation attempt activates the same predefined set of template positions. The active-node indicator a remains fixed, and the available device/port positions are independent of the task.

Within this template, the generator preserves Grid and predicts Routing and Bridging to determine each candidate circuit’s connectivity. All baselines receive the same device/template information and share the same retrieval library and retrieval permissions.

R1.3. Splitting precedes augmentation, and retrieval-library updates are disabled during evaluation

Data are partitioned by original circuit ID before augmentation. Samples produced through device substitutions, specification variations, and text paraphrases inherit their source circuit’s split assignment. All augmented samples derived from the same original circuit therefore remain in the same split. Training circuits come from AnalogGenie (Appendix B), while the two evaluation benchmarks are derived from OCB and AMSBench (Appendix D).

The retrieval library is built from a separately prepared circuit collection and separated from test data before evaluation. The successful-design storage operation in Figure 3 and Algorithm 2 is intended for day-to-day use and is disabled during evaluation. Designs produced by test tasks therefore do not enter the retrieval library for subsequent test tasks.

R1.4. Shared evaluation settings and the contribution of the multi-layer representation

All methods in Table 2 use the same tools, inputs, and budget limits. Every baseline, including AnalogSeeker, uses the same inner sizing optimizer (Algorithm 1), device/template information, and retrieval permissions. The budgets are K=15, T=10, and at most 400 Ngspice calls per candidate. Each method stops at the first independently verified feasible design; unsuccessful tasks exhaust 15 topology attempts. The per-candidate budget covers initialization, perturbations, trial evaluations, failed calls, and final verification (Appendix C).

The multi-layer representation improves success rates by 35 and 25 percentage points on the two Hard splits. The full representation achieves 66% / 51%, compared with 31% / 26% for the single-layer variant. The single-layer variant uses the same template and node indexing, encoding connections in a single layer. Grid, a, the generator backbone, training data and settings, downstream pipeline, and budgets remain unchanged, as do the matrix-to-netlist decoding rules. These results support explicitly separating connection types through a multi-layer representation.

ANoA is defined as the average number of outer topology attempts over successful tasks, excluding inner parameter updates. Runtime is also computed over successful tasks. Evaluation is conducted on a platform equipped with a single NVIDIA H100 GPU.

We thank the authors for the detailed response.

The matched 15-attempt comparison is the control we asked for. Feedback adds 8.5 and 8 points; most of the gain over the single-attempt ablation comes from additional attempts. Please include this table in the paper with some measure of uncertainty.

The shared template and common budgets for all baselines address our fairness concern.

Splitting by circuit ID and disabling retrieval-library updates help. A topology-level overlap check between AnalogGenie and the test sets would still be useful.

Most of our concerns are addressed.

Thank you for your constructive follow-up and for confirming that most of your concerns have been addressed. We will include the matched 15-attempt comparison in the revised manuscript with uncertainty estimates, and explicitly report the topology-level overlap check between the AnalogGenie-derived training data and both evaluation benchmarks.

In light of the additional experimental evidence and protocol clarifications, we would be grateful if you could reconsider your overall assessment and recommendation. We would be happy to address any remaining questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.