acceptodds
Withdrawn from ICLR 2027

CIRCUIT-R1: CIRCUIT-GROUNDED REINFORCEMENT LEARNING FOR VERIFIABLE CIRCUIT2NETLIST

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent multimodal large language models have shown strong visual reasoning capabilities, yet reliable circuit-to-netlist translation remains challenging. Exist- ing circuit-oriented reasoning methods use reinforcement learning to reward only a subset of intermediate reasoning stages, leaving the grounding of circuit con- nections insufficiently constrained. We investigate this limitation from two com- plementary perspectives. From an explicit perspective, our analysis shows that recognizing devices and ports alone is insufficient: wire tracing provides essential geometric evidence for recovering circuit topology. From an implicit perspec- tive, generated netlists frequently violate basic structural and electrical principles, demonstrating the need for circuit-specific priors beyond visible image evidence. Motivated by these findings, we introduce CIRCUIT-R1 (Circuit-Grounded Re- inforcement Learning for Verifiable Circuit-to-Netlist Translation) . CIRCUIT-R1 decomposes netlist generation into five reasoning stages and defines continuous reward signals over bounding boxes, terminal centers, wire polylines, and end- point assignments. We further propose the Circuit Physics Prior Verifier (CPPV), which represents each predicted netlist as a typed terminal hypergraph and detects structural and electrical violations without consulting the ground-truth netlist. We will release Cir-GCoT, a dataset of 2,440 circuit schematics with fine-grained grounding and topology annotations. Our results show that explicit visual evi- dence and implicit circuit priors jointly guide policy optimization toward accurate and reliable netlist generation.

Primary area:
reinforcement learning
Keywords:
Circuit-to-Netlist Translation, Multimodal Large Language Models, Visually Grounded Reasoning, Circuit-Physics Verification

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_1fmAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper did some specific improvements on the task of transforming analog circuit images to SPICE format with VLLM. Compared with existing works like Circuit-Think, the major improvements contains:

  • The CoT information is extended from 3 steps to 5 during SFT. Device ports and wire tracing information are added
  • The reward is more calibrated during RL, including polyline F1 and evidence binding reward
  • A verifier (named CPPV in the paper) that penalize circuits that break basic physical rules (e.g., missing terminals, internal shorts) during RL
Critical strengths
  • A verifier based on physical rules makes very good sense in the circuit2netlist task
Critical weaknesses
  • This paper seems to be a regular iteration upon existing works like Circuit-Think, extending and crafting CoTs and rewards.
  • The performance improvement is also pretty "regular". I won't say it is marginal, but considering that circuit2netlist is a very application-oriented task, and this paper focuses solely on this task, I can’t help but wonder about the practical significance of this article. Sadly, in real application, the difference between 66% and 76% accuracy may not be as large as it seems (from "not reliable at all" to "a bit better, but still not reliable at all"). To make matters worse, the metric of accuracy here seems already a pretty relaxed one, allowing partial correct circuits to gain scores. I wonder how would a "netlist exact match rate" looks like, i.e., the precentage of SPICE netlist that exactly recovered the circuit image.
Questions for authors

See "weaknesses"

Reviewer_K7VZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies circuit-to-netlist translation with a vision-language model. A motivating analysis on AMSNet argues two things. First, frontier models (Gemini-3.1-Pro, Kimi-K3, GPT-5.5) often get every device and port right but still produce the wrong topology (48.9 to 63.5% of entity-correct predictions). Second, a stage-wise oracle on a cold-start SFT model shows the largest jump (+19.57 points, 75.82 to 95.39) when ground-truth grouped wire polylines are supplied. The authors also show that 29 to 35% of frontier-model netlists violate simple structural rules, and that 7.6 to 12.1% still do among predictions with Answer Accuracy of at least 0.8.

The method, Circuit-R1, fine-tunes Qwen3-VL-4B with SFT on 1,000 samples and then GRPO on 2,014 samples. Training uses a five-stage textual reasoning trace: external ports, devices, device-port coordinates, wire polylines, and endpoint groups. The rewards are continuous IoU for ports and devices, distance-based terminal rewards, point-set precision and recall for polylines, a Hungarian-matched binding reward between traces and endpoint sets, and a net-multiset Jaccard answer reward. A reference-free penalty (CPPV) counts missing terminals, singleton nets, undeclared endpoints, duplicates, and intra-device shorts. The model reaches 76.60 Answer Accuracy on the authors’ new Cir-GCoT test split (426 samples) and 77.67 on AMSNet, against 62.83 and 75.07 for the strongest frontier baseline. The authors also report that adding the generated netlist to the prompt improves three VLMs on five AMSBench tasks.

Critical strengths

S1. Beats the competing models

S2. Promises to open-source a dataset

Critical weaknesses

W1. The contribution is narrow, with no algorithmic novelty beyond reward design. The optimizer is unmodified GRPO, the backbone and SFT stage are standard.

W2. The paper has no related-work section, and its relationship to Circuit-Think (AAAI 2026), its direct predecessor, is unclear.

W3. Relevant prior work is missing, and some citations are incorrect. The paper omits Masala-CHAI (the current version of the cited Auto-SPICE preprint), SINA, Netlistify (MLCAD 2025), and Image2Net for image-to-netlist conversion, as well as Visual-RFT, which already uses continuous IoU rewards under GRPO. Two of the three citations given for computer-vision circuit-parsing pipelines are unrelated to the task. Zhang et al. 2022 is an image compressive sensing network, and Jin et al. 2025 addresses computational lithography. The same incorrect citations appear in the corresponding passage of Circuit-Think, which suggests the references were carried over without verification.

W4. The Circuit-Think baseline is reported inconsistently. The footnote to Table 2 states that Circuit-Think was reproduced because its data and code are unavailable. However, the AMSNet number (69.32) is identical to the result reported in the Circuit-Think paper, and this is not disclosed. The Cir-GCoT number (58.54) is identical to the thresholded-IoU variant in Table 10, which appears to be the authors’ own pipeline with one reward changed. Yet the row is labeled 7B while the authors’ backbone is 4B. The row therefore seems to combine a copied number and an in-house ablation under one name.

W5. There is no SFT-only or plain-GRPO baseline, so the paper never shows that its reward design is needed. The only SFT number appears incidentally as condition C0 of the oracle analysis on AMSNet (57.63), and no SFT result is given on Cir-GCoT. The SFT stage also uses only 1,000 samples while RL uses 2,014, so a fair SFT baseline should be trained on all 2,014 annotated traces.

W6. The headline metric does not measure correctness, and the model is trained directly on it. Answer Accuracy is an exact-net multiset Jaccard score (Eq. 14), and R_answer is the same function. The metric gives substantial credit to wrong netlists, and Figure 4 shows this directly.

Questions for authors

Q1. Was the Circuit-Think AMSNet result (69.32) taken from the Circuit-Think paper? Is the Cir-GCoT result (58.54) the same run as the thresholded-IoU variant in Table 10, and which backbone does it use? How does the claim that the data is unavailable square with the dataset repository listed in the Circuit-Think paper?

Q2. Please state precisely which components of the reward and training pipeline differ from Circuit-Think. Why were the graph consistency reward and reflective learning removed?

Q3. Can you train Circuit-R1 on the Circuit-Think dataset, or the Circuit-Think reward on Cir-GCoT, so that method and data effects can be separated?

Q4. Please report SFT-only results trained on all 2,014 annotated samples, and GRPO with only the answer and format rewards, on both benchmarks.

Q5. Please report exact netlist accuracy for all methods in Table 2, stratified by circuit size.

Q6. What are the sources for the 1.5T and 2.8T parameter counts? How do frontier models perform with thinking enabled?

Q7. Were the reward weights selected on the test split? Please report results with weights chosen on a validation split, with means and standard deviations over several seeds.

Q8. Where do the Cir-GCoT schematics come from, and what is their overlap with AMSNet, AMSBench, and the Circuit-Think dataset?

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

This work focus on circuit-to-netlist translation problem. It makes an analysis to show that recognizing devices and ports alone is not sufficient. Motivated by it, it introduces circuit-R1, which decomposes netlist generation into give reasoning stages and defines reward signals over bounding boxes, terminal centers, wire polylines, and endpoint assignments. It further introduces circuit physics prior verifier to detect structural and electrical violations.

Critical strengths
  • The paper is well-written and easy-to-follow.
  • The motivation behind the proposed approach is well explained by making interesting and insightful analysis.
  • The experimental results are comprehensive, demonstrating the effectiveness of the proposed approach.
Critical weaknesses
  • What is the source of Cir-GCoT dataset? Specifically, where does the input and output come from, e.g., a existing dataset or collected by the authors? Have it went through quality filtering? Does it have overlap with AMSNet? And, how is the reasoning collected?
  • For potential hyper-parameters in reward and verifier (section 3.1.1 and 3.1.2), could you give a full list to summarize all hyper-parameters? And how do these hyper-parameters influence the performance of the proposed approach?
  • The SFT and RL use the same set of training data. Will it achieves better performance if RL uses a different and harder set?
Questions for authors

See weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.