acceptodds
Withdrawn from ICLR 2027

AnalogQuad: Learning Analog Circuit Representations through Template–Instance Neural-Symbolic Alignment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Analog circuit representations should support reuse across designs while retaining the behavior induced by device sizing and bias. Graph-centric pretraining captures topology, but it does not directly expose a structured small-signal symbolic behavior view that organizes gain, poles, zeros, and transfer functions; global graph–symbolic alignment may recognize a circuit family without resolving its different realizations. Unlike ordinary paired modalities, a parameterized symbolic record is instantiated with the sizing parameters in the netlist. We represent graph and symbolic views at template and sized-instance levels, , with a sizing assignment linking the two levels. We introduce AnalogQuad, which combines sizing injection into graph node features and symbolic records with shared template–instance encoders and a hierarchical symbolic tower that preserves record-entry structure. A symmetric contrastive objective aligns graph–symbolic pairs through either sequential training on the two domains or joint training on mixed-domain batches. Experimental results demonstrate that template-level pretraining alone has limited transferability, whereas subsequent adaptation to sized instances consistently improves performance on downstream tasks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
analog circuits, representation learning, neural-symbolic learning, multimodal learning, circuit representation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aLVPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AnalogQuad aligns circuit graphs with symbolic small-signal records at template and sized-instance levels using sizing injection, shared encoders, and contrastive learning. Evaluation covers retrieval, functional annotation, and performance classification. Joint training achieves bidirectional MRRs of 0.841 on templates and 0.113 on sized instances, with modest downstream improvements over training from scratch.

Critical strengths
  1. Symbolic supervision provides a well-motivated connection between circuit structure and parameter-dependent behavior. Joint training preserves template retrieval while improving instance retrieval.

  2. Evaluation spans four graph backbones and five downstream seeds, with transparent reporting of cases where GraphCL performs better.

Critical weaknesses
  1. Component benefits are insufficiently isolated. Pretraining budgets differ, sizing-only pretraining is not evaluated.

  2. Generalization evidence is limited: sizing evaluation contains only two test topologies, and cross-domain topology overlap is unspecified.

Questions for authors
  1. Do template-training topologies overlap sizing validation/test topologies or downstream test circuits?

  2. Do instance-level gains persist against compute-matched sizing-only pretraining and when retrieval negatives share the query’s topology?

Reviewer_Vj86OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a learning framework that aligns analog circuit graphs with symbolic descriptions. The manuscript seems to be largely written by AI and lacks a necessary explanation of the motivation. I have a major concern about the method's motivation.

Specifically, the author mentions: "A key challenge is to learn reusable representations that support retrieval, functional annotation, and design analysis across circuits with related functions." I have difficulty understanding the difference between the reusable representation and the singlerepresentation from an encoder. Additionally, I can not identify the importance of the reusable feature of a representation, and what the obstacle is in this direction. In the paper, the main application scenario is functionality-aware retrieval. The main challenge seems to be deciding which similarities and differences the representation should preserve. Additionally, I don't understand why the symbolic supervision makes a representation reusable.

The central contribution of AnalogQuad is the symbolic supervision. But I can not find supporting evidence in the experiments that such symbolic learning provides useful information beyond what the circuit topology and device-size features already provide.

This paper seems to simply pick hot topics like graph learning and symbolic learning and try to apply them to analog circuit design. I think symbolic learning for circuit design is promising, and I appreciate the authors' efforts, but they should also carefully consider whether the motivation is sound.

Critical strengths

The method uses symbolic supervision in graph learning to provide additional information for circuit retrieval. This is a novel approach.

Critical weaknesses
  1. The writing quality of the paper is poor. The paper lacks enough explanation of the motivation. The statement that reusable representations is a key challenge is questionable to me. The proposal to use symbolic supervision for graph learning does not seem to connect to the reusable representationsin a necessary way. The authors seem to heavily rely on AI to generate the manuscript, which makes it significantly hard for me to follow and investigate the contribution of the work.
  2. The experiments do not provide enough evidence to support the claim of contributions in the paper. The improvement from the symbolic learning is not demonstrated effectively.
Questions for authors
  1. I would appreciate it if you could provide more background information on how circuit retrieval is applied and why it is important in analog design. Additionally, I want a clear definition and explanation of reusable representations and why it is important in this field. What is the difference between the representation and reusable representation? I also want more explanation on the challenges of obtaining such reusable representations.
  2. Can you provide more explanation of the technique details on how to connect symbolic supervision with the reusable features? Additionally, can you provide more experimental results to demonstrate the effectiveness of symbolic learning? For example, how does the symbolic learning provides more information than the existing graph structure and size information?
Reviewer_UJzhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AnalogQuad, which learns analog circuit representations by aligning circuit graphs with small-signal symbolic records at both template and sized-instance levels. The graph and symbolic encoders share parameters across the two domains, while a hierarchical symbolic encoder combines expressions with entry metadata. Experiments evaluate cross-modal retrieval and transfer to functional annotation and performance classification. Joint training achieves the best retrieval results among the compared configurations, while downstream gains from sequential pretraining vary across tasks and graph encoders.

Critical strengths
  1. Using small-signal symbolic descriptions to complement circuit connectivity is a meaningful direction for learning representations that reflect parameter-dependent behavior.

  2. The experiments cover both retrieval and downstream transfer, with improvements across several graph encoders and evaluation settings.

Critical weaknesses
  1. The independent contribution of the hierarchical symbolic encoder remains unclear. In Table 2, flat BERT has a higher mean retrieval score, but the authors acknowledge that input content, aggregation, and template pretraining history also differ. This makes it difficult to attribute the results to the encoder design itself. MM also uses a hierarchical encoder, and its training settings differ from those of AnalogQuad, so comparisons with MM might not resolve this issue.

  2. The overall novelty of this work seems incremental. Contrastive learning is already widely used. This work seems to apply the general strategy from other domains to graph–symbolic alignment. While this application to the analog circuit is meaningful, the methodological contribution may not be that significant. I desire to see a discussion in terms of technological innovations, specifically designed for this field.

Questions for authors
  1. I find the comparison between the flat and hierarchical encoders a little difficult to interpret, since the inputs and pretraining history also differ. Would it be possible to compare them using the same records and training setup?

  2. With the encoders shared across domains, I am wondering whether a test topology in one domain could have appeared in the other domain’s training set. It would also be helpful to see retrieval within each of the two sizing test topologies, using only candidates from that topology, to better understand how well the model distinguishes different sizing instances.

  3. How does AnalogQuad relate to CktGen [1] and Ckt2Vec [2]? CktGen aligns circuits with performance specifications using contrastive learning for analog circuit generation, while Ckt2Vec incorporates device electrical characteristics into circuit representations. A brief discussion would help clarify what additional information symbolic records provide and what learning at both template and sized-instance levels contributes.

[1] CktGen: Automated Analog Circuit Design with Generative Artificial Intelligence

[2] Ckt2Vec: Efficient Electrical Encoding for Analog Circuit Representations in Vector Space

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

Related papers

Discussion (0)

Sign in to comment.