acceptodds
Under review as a conference paper at ICLR 2027

ShareGate: Learning to Match Functional Subgraph With Shared Logic

Abstract

Functional subgraph matching identifies circuit regions that implement the same Boolean functions as query patterns, supporting datapath optimization, arithmetic verification, and hardware Trojan analysis. Logic synthesis creates shared intermediate expressions, so functional regions overlap and gates serve multiple operations. This shared logic interferes with learned recognition: neighborhood aggregation mixes operation cues at shared gates, leaving gate classes ambiguous. To address this, we propose ShareGate, a framework that recovers and certifies arithmetic cores in post-synthesis netlists with shared logic. Within ShareGate, a graph neural network proposes a candidate region for each gate from per-gate descriptors. Verification-driven refinement accepts a region only when its simulated response reproduces a fixed reference model on every pattern in the budget. A scheduler orders the checks. Thus, a class label changes only through a passing functional test, and the accepted region carries that test as a certificate. We formalize the task as functional subgraph matching with shared logic. Each gate is assigned one of five classes (adder, multiplier, subtractor, comparator, and control), so arithmetic regions are labeled together with their embedded comparison and control logic. Experiments on 21 netlists with embedded arithmetic cores from EPFL, ISCAS-85, and ITC-99 show that ShareGate recovers and certifies all three arithmetic cores in every case at all four injection levels, reaching certified results 6.0 to 34.5 times faster than the fastest baseline. As a diagnostic of the recognition stage, the gate-level field reaches 92.3% accuracy at the deepest level, versus 76.8% for the strongest baseline.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Subgraph matching, circuit verification, logic synthesis

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_UdJfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes a GNN learning framework to predict a subgraph with shared logic of a post-synthesis netlist. This is generally an interesting direction, and I have seen related work in the community. But I have several concerns about the problem formulation and the application's meaning.

Critical strengths

Prediction of functional subgraphs with shared logic is a promising research direction in EDA. I appreciate the authors' efforts to utilize GNNs to address the challenging prediction problem.

Critical weaknesses
  1. The problem formulation is unclear. I have a big concern about the meaning and effectiveness of the method. I found the problem formulation scattered throughout the paper. As a result, I find it hard to understand the task setting. For example, is the input a single netlist or a pair {netlist, RTL}? The paper formulates the task as a gate-level annotation task, so how it deals with the scenario where one gate is shared with several different functional subgraphs remains unclear. The unclear problem formulation also raises a significant concern about the meaning of such an application, because commercial EDA tools already support substantial parts of similar requirements. For example: Synopsys Formality: RTL–netlist correlation, equivalent-net matching, and synthesis guidance covering transformations such as multiplier architectures Cadence Conformal supports RTL-to-gate equivalence checking, including complex arithmetic logic and datapaths. Synopsys VC Formal Datapath Validation supports arithmetic datapaths and reference-model-based checking. It would be helpful if the author could clearly distinguish between the tasks resolved by commercial tools and the remaining gap that necessitates GNN-based prediction. GNN-based prediction cannot guarantee 100% accuracy, which always raises concerns about the reliability of circuit design. As a result, if such applications can largely be resolved with off-the-shelf commercial formal verification tools, the value of GNN-based methodology will be significantly weakened.

  2. The contribution and novelty of the method are not clearly introduced. The authors list three contributions at the end of the introduction: (1) a problem formulation, (2) a network with simulation-based checks, and (3) evaluation on public benchmarks. Contribution (1) is problematic since it conflicts with the scenario where one gate is shared across different functional subgraphs. The most important contribution should be around (2), the training framework, which combines recognition with verification-driven refinement. What is verification-driven refinement? Why does such a VDR process contribute to better prediction performance?

  3. The writing quality is low. The manuscript appears to be mainly AI-generated, and many presentations are ambiguous and hard to understand, which significantly hinders an effective evaluation of the paper's contributions.

Questions for authors
  1. I would appreciate a clear introduction to the background, including how this problem is addressed by commercial verification tools, and why it still remains challenging and necessitates GNN-based learning methods.
  2. Can you clarify the problem formulation, especially what the input to the model is? Is the input a single netlist or a pair {netlist, RTL}?
  3. Can you clarify what '' verification-driven refinement '' is? Why does this VDR process improve prediction performance? Is there any challenge that prevents the general GNN-based learning method from working for the task?
4 comments
AuthorsOpenReview ↗

Response to Reviewer UdJf

We thank the reviewer for seeing functional subgraph prediction under shared logic as a promising direction in EDA. The revision states the problem formulation in one place (Section 2.3) and positions the work against commercial equivalence checking (Section 2.2 and Appendix O).

[W1, Q1] How do commercial tools address this problem, and why is a GNN needed?

Response.Commercial tools solve a fundamentally different problem and cannot replace region identification. Tools like Formality, Conformal, and VC Formal are equivalence checkers that verify whether a netlist matches its RTL. They cannot answer "which gates implement which operation?", nor can they run without RTL, such as for third-party IP or legacy designs. Even with RTL, overlapping logic remains highly intractable for them. When synthesis merges operations of several bit widths into a datapath, the internal correspondence points disappear, often causing equivalence checking on such circuits to fail. Knowing the region of each operation is exactly what makes these checks tractable, which is the goal of our task.

A GNN is needed because traditional template matching fails after synthesis. Template matching relies on fixed stored patterns, but synthesis generates countless structurally different implementations for the same operation. Furthermore, shared logic breaks this approach entirely: a shared gate belongs to two operations at once, so neither operation appears as a complete structural match. A GNN solves this by learning from synthesized examples to infer a gate's function from its context, making it robust to both structural changes and overlapping logic. We have added this background to Sections 1 and 2.2..

[W1, Q2] Is the input a single netlist or a pair {netlist, RTL}?

Response. The input is a single gate-level netlist. No RTL is needed. ShareGate also uses a fixed library of generic reference functions (adder, subtractor, multiplier, comparator), which is the same for every netlist.

The output is one class for each gate, together with the region of every certified arithmetic core. A gate shared by several operations receives one class label and is included in the region of every core that uses it, so sharing is represented by the regions. We have gathered the formulation in Section 2.3 and added the input, the output and the treatment of shared gates to Definition 2.

[W2, Q3] What is verification-driven refinement, why does it improve prediction, and what prevents a general GNN from working?

VDR verifies each region proposed by the GNN through simulation, retaining a label only if the region passes the check. It applies the same 64 random input patterns to both the proposed region and a reference function—for instance, a 40-bit adder. If the outputs match for every pattern, the gates within that region are assigned the corresponding class; otherwise, the proposal is either revised or discarded. The complete, closed regions generated by VDR can ultimately be fed into a SAT/Miter solver for 100% rigorous formal equivalence verification (a process successfully completed for all 84 out of 84 cases, as detailed in Appendix K.7).

A general GNN faces two obstacles on this task. First, it labels a gate from its neighbors, and the neighbors of a shared gate belong to two operations, so the prediction there is ambiguous. Second, per-gate labels do not form a complete region that can be verified: regions taken directly from the classifier output are certified in 0 cases, against 252 of 252 after VDR. VDR improves prediction because it corrects labels with the function of a whole region, which the GNN cannot see. On the same classifier output it adds 20.9 to 32.3 MCC points over label smoothing (Table 2). This also answers the reliability concern: a GNN prediction is only a proposal, and no region is reported without passing a functional check.

[W3] Writing quality.

We have rewritten the introduction, Section 2.3 and Section 3.3 for clarity.

AuthorsOpenReview ↗

Follow-up on task formulation and verification-driven refinement

Dear Reviewer UdJf,

Thank you again for highlighting the need for a clearer task formulation and explanation of verification-driven refinement. We would appreciate your feedback on our November 15 response before discussion closes on November 18.

Our response explains the single-netlist input, overlapping region membership, and the distinction between region identification and equivalence checking. It also describes simulation-based refinement and the separate formal equivalence checks, with pointers to Sections 2.2, 2.3, and 3.3.

Please let us know of any remaining concerns about the task setting or verification procedure so we can clarify them. If the response and revision resolve your main concerns, we would appreciate an updated assessment and rating reflecting these clarifications.

Thank you for your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

Gentle follow-up before the discussion closes

Dear Reviewer UdJf,

Thank you again for your time and careful review. With the discussion period ending on November 18, we would be grateful for your feedback on our November 15 response.

Our response covers the single-netlist input, shared-gate membership, and the role of verification-driven refinement. Please let us know of any remaining concerns so we can clarify them before discussion closes.

We appreciate your consideration of the response and revised manuscript in your final assessment.

Best regards, The Authors

Reviewer_UdJfOpenReview ↗

Response

Dear authors, Thanks very much for your response to my questions. I have noticed the modifications to the paper, which actually provide meaningful clarifications, and I am able to understand your tasks and formulations better. However, the modifications to the original paper seem excessive and almost rewrite the introduction and motivation sections. I think this might be a concern, as it substantially rewrites a majority of the original manuscript. But considering the author's efforts, I decided to raise my score to 3 (weak accept).

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

This paper proposes ShareGate, a framework that combines graph neural network-based gate classification with verification-driven refinement to identify and functionally verify arithmetic cores in circuits containing shared logic. Logic synthesis can cause different arithmetic cores to share gates, making it difficult to identify functional components such as adders, multipliers, and subtractors within larger circuits. The proposed framework addresses this challenge by generating candidate regions, checking them against reference functions using simulation, and refining their boundaries and gate labels based on the results.

Critical strengths
  • The paper addresses an interesting and relevant problem in circuit analysis.
  • The proposed framework can serve as a foundation for future work combining learning with formal verification in this area.
  • ShareGate consistently outperforms the nine evaluated baselines across all four classification metrics, achieving 92.34% accuracy compared with 76.80% for the strongest baseline at the highest injection level. Furthermore, ShareGate demonstrates substantial improvements in efficiency, achieving time speedups of at least 6x the baseline.
Critical weaknesses

The paper may want to address the following aspects in a revised version:

  • How realistic are the simulated injections used to cause interference?
  • Would ShareGate work well on arithmetic circuits of different sizes and designs?
  • Equation (4) appears inconsistent with Proposition 1 and Appendix C.4, as it updates all gates in an accepted cone rather than restricting updates to previously unlocked gates. This should be clarified.
  • Revise for typos, e.g., space after the period in line 50.
Questions for authors

See questions listed above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer bHMB

We thank the reviewer for the positive assessment and for viewing ShareGate as a foundation for combining learning with formal verification in this area. The four points are answered below.

[W1] How realistic are the simulated injections?

Response.Observation gates realistically simulate actual shared gates. While actual shared gates account for only 2.70% to 4.23% of the gates across the 21 evaluation netlists at ρ = 0, even this small proportion makes identifying functional regions challenging.

Physically, observation gates reproduce the effect of a shared gate: two supports cross at one gate (Definition 5), so α̃ grows with ρ while κ stays unchanged. Behaviorally, as observation gates increase, the mean raw accuracy of the nine baselines falls monotonically. Since synthesis cannot precisely set the proportion of actual shared gates, we use observation gates in the main experiments to control the injection dose exactly.

To directly evaluate actual sharing, we built the synthesis-variant panel (as detailed in our response to Reviewer nvii, Q1–Q2). On the same 21 netlists, actual shared gates rise to 22.9%, 40.2%, and 53.4% of the gates. The baseline trend matches the observation-gate sweep: raw accuracy on shared gates falls as sharing grows for eight of the nine baselines (Table R1). Under both perturbations, VDR improves every model and ShareGate leads. On actual sharing, ShareGate reaches 99.74%, 99.44%, and 98.96% on shared gates and certifies every adder and subtractor core at every level (Table R2).

Table R1. Accuracy (%) on shared gates (κ ≥ 2), raw → +VDR, mean over the 21 netlists. Column heads give the mean share of shared gates. (Manuscript Table 15.)

ModelShared gates 22.9%Shared gates 40.2%Shared gates 53.4%
ShareGate (ours)98.79→99.7490.17→99.4489.62→98.96
ABGNN95.46→96.6186.58→94.8882.19→90.81
DAGNN-RE87.46→88.1872.92→83.9658.48→71.26
GAT90.49→93.3181.75→91.5376.70→84.82
GCN77.60→89.5666.90→84.3455.37→71.88
GraphSAGE85.24→88.2671.96→84.1758.61→71.65
HGCN66.76→95.8974.50→91.9672.99→85.55
NeuroMatch67.66→87.4539.08→78.9128.36→62.08
ReVEAL59.48→89.2945.51→79.2339.76→62.70
WideGate92.03→96.6287.81→95.4984.52→91.95

Table R2. Certification by VDR. (Manuscript Table 16.)

SettingADD cores certifiedSUB cores certified
No sharing (L0)21/2121/21
~0.23 (22.9%)21/2121/21
~0.46 (40.2%)21/2121/21
~0.73 (53.4%)21/2121/21

[W2] Would ShareGate work on arithmetic circuits of different sizes and designs?

Response. ShareGate generalizes robustly across different circuit sizes and designs. Even when tested on configurations outside its training distribution (widths of 12 to 40 bits vs. training widths of 8 and 32), ShareGate consistently achieves high accuracy. Specifically, at ρ = 0, it reaches at least 96.95% Acc and 92.31 MCC across all 12 configurations (1 to 8 cores, four functional mixes, and six bit widths from 12 to 40). At ρ = 0.5, Acc ranges from 83.18% to 99.29%, with the lowest value occurring at the narrowest width (12 bits). These configurations include Priority, c2670, and b10. Table R10 and manuscript Table 20 report all metrics.

Table R10. ShareGate across core count, functional mix and bit width (Acc / MCC, %).

Configurationρ = 0ρ = 0.5
1 core (ADD)96.95 / 92.3186.92 / 78.43
2 cores (ADD, MUL)99.48 / 98.4498.42 / 95.28
4 cores (ADD, MUL, SUB, CMP), 40 bits98.93 / 97.1494.42 / 85.43
8 cores ((ADD, MUL, SUB, CMP) × 2)99.58 / 98.7994.79 / 84.34
Mix ADD × 498.62 / 97.3092.97 / 87.81
Mix ADD, MUL, ADD, MUL99.83 / 99.2399.29 / 96.82
Mix MUL, MUL, SUB, CMP99.68 / 98.5696.88 / 83.04
4 cores, 12 bits97.07 / 95.0883.18 / 75.55
4 cores, 20 bits98.83 / 98.2388.52 / 81.85
4 cores, 24 bits98.46 / 97.5390.02 / 83.02
4 cores, 28 bits99.11 / 98.4491.19 / 83.15
4 cores, 36 bits99.40 / 98.6794.07 / 85.85

Q3. [W3] Equation (4) is inconsistent with Proposition 1 and Appendix C.4.

Q4. Revise for typos, e.g., space after the period in line 50.

Response. According to these suggestions, we have made revisions in the latest manuscript.

AuthorsOpenReview ↗

Follow-up on actual sharing and circuit variations

Dear Reviewer bHMB,

Thank you again for your constructive feedback and positive assessment of the work. As the discussion period ends on November 18, we would appreciate your feedback on our November 15 response.

For ease of reference, Tables R1-R2 cover the additional actual-sharing experiments, and Table R10 covers variations in circuit size and design. The final part of our response addresses Equation (4) and the presentation corrections you requested.

Please let us know of any remaining concerns; we would be happy to clarify them before discussion closes. We would also appreciate your assessment of the additional evidence in your final review.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_bHMBOpenReview ↗

After rebuttal

Many thanks to the authors for addressing my concerns. After reading all the reviews, I stand by my original score, which is already leaning toward acceptance.

Reviewer_nviiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

ShareGate addresses functional subgraph matching in post-synthesis netlists where logic sharing lets a single gate serve several arithmetic operations, confusing GNN classifiers that aggregate over neighborhoods. It pairs a fixed GraphSAINT classifier that proposes a five-class gate label field (adder, multiplier, subtractor, comparator, control) with verification-driven refinement: candidate regions are simulated against fixed reference models under a UCB1-scheduled, bounded check budget, and only a window that reproduces the reference on all 64 patterns can rewrite labels in its support cone, which is then locked, so every label revision carries a functional certificate. On 21 EPFL/ISCAS-85/ITC-99 netlists with embedded 40-bit adder, 40-bit subtractor, and 20-bit multiplier cores under four levels of injected observation gates, ShareGate reaches 92.3% gate accuracy at the deepest level versus 76.8% for the best of nine GNN baselines, certifies all three cores in every case, and reaches a certified result 6.0–34.5× faster than the fastest baseline on a five-netlist timing panel, with ablations showing that label smoothing on the same prior gains at most a few MCC points while refinement gains 20–32.

Critical strengths
  1. The core design decision is the right one for this problem, and the paper demonstrates it cleanly: a learned classifier only proposes, and a simulation conformance check against a reference model is the sole mechanism that can change a label, with accepted cones locked so later rounds cannot undo a certified assignment. Table 2 is the evidence that matters: three post-processors share the same prior, label voting and MRF smoothing move MCC by 2–4 points, and refinement moves it by 20–32 at every level. That isolates the contribution to the decision rule, and it is why I take the method seriously despite the evaluation’s problems.
  2. The paper connects gate-level recognition to a certificate, which prior GNN work in this space does not do. Appendix K.7 shows that regions induced directly from the classifier’s field certify on zero cases because they form open cuts, while refined regions certify on 252/252 core attempts and pass the plain-miter equivalence check on 84/84. Whatever one thinks of the benchmark construction, turning a label field into a boundary that a SAT solver accepts is the output that downstream verification tools actually need, and the paper is the only one of its baselines to produce it.
Critical weaknesses
  1. The sweep does not vary shared logic. Section 4.2 states that the appended observation gates carry κ = 0 and leave α unchanged; they only add label-free neighbors. The title, abstract, and headline claims (“all four injection levels,” “more robust to increasing logic sharing”) therefore rest on a variable that measures neighborhood dilution, while the only real sharing is whatever synthesis produced at ρ = 0, which is never quantified. The central empirical claim is about something the experiment does not manipulate.
  2. The speed and certification comparison is against an author-built post-processor, not the published methods. Every baseline is run through the authors’ “toolbox” traversal and lands at 130–150 s regardless of field quality, which means the 6–34.5× margin is refinement versus that traversal. ShareGate’s own classifier is also weaker than ABGNN and DAGNN-RE at ρ = 0 (83.8% vs. 95.9%), yet refinement is never applied to a baseline field, so the paper cannot say whether its gain comes from the method or from the fact that only one arm received it.
Questions for authors
  1. How much actual shared logic (gates with κ ≥ 2, or α ≥ 2) do the evaluation netlists contain at ρ = 0, per netlist? If the answer is small, the title claim needs to be re-scoped; if it is substantial, please report gate accuracy and certification restricted to those gates.
  2. Can you run a sweep that varies sharing rather than dilution, e.g., by varying comparator width or the number of operations that reuse an intermediate expression? That is the experiment the abstract describes.
  3. What exactly does the baseline “toolbox” do, and why do all nine baselines take 130–150 s independent of their label quality? Please report the time each baseline needs when given the same proposal-and-check refinement that ShareGate uses, so the comparison is method-to-method.
  4. Please apply your refinement stage on top of the ABGNN and DAGNN-RE fields. Their classifiers outperform the proposed method at ρ = 0 (95.9% and 94.5% vs. 83.8%), so if refinement works on any prior, the paper’s contribution is refinement alone and the classifier is irrelevant.
6 comments
AuthorsOpenReview ↗

Response to Reviewer nvii (1/3) - Q1-Q2

We thank the reviewer for the careful review. Our answers follow.

[W1, Q1] How much actual shared logic do the evaluation netlists contain at ρ = 0?

Response. At ρ = 0, actual shared gates (κ ≥ 2) account for 2.70%–4.23% of gates across the 21 evaluation netlists. Even with so few actual shared gates, identifying functional regions remains challenging. In Definition 1, readers of a shared gate v (κ(v) ≥ 2) touch two supports at once (α̃ ≥ 2), so neighborhood aggregation mixes the cues of two operations and the classifier fails. An observation gate is a fan-out node formed by the intersection of two gates; it reads one gate from each region, making the two gates two-step neighbors in the undirected neighborhood and producing the same α̃ mixture. It is therefore an alternative way to simulate the impact of shared gates. Because synthesis cannot precisely set the proportion of actual shared gates, the main experiment uses observation gates to control the dose exactly (see the revised Section 4.2). We have re-scoped the wording: the abstract and Sections 4.1–4.3 now call ρ the observation-gate injection level. We further investigate ShareGate with higher levels of actual shared gates. On the same 21 netlists, we implement the adder, subtractor and comparator so that they reuse a common intermediate expression (setup in [W1, Q2]). Synthesis merges the common logic, and actual shared gates rise to 22.9%, 40.2% and 53.4% of gates. As requested, Table R1 reports accuracy restricted to actual shared gates. ShareGate is the most accurate on actual shared gates at every level, both before and after refinement, and every adder and subtractor core is certified at every level (21/21 netlists each). The comparator has a single-bit output and is not a certification target (Section 3.3).

Table R1. Accuracy (%) on shared gates (κ ≥ 2), raw → +VDR, mean over the 21 netlists. Column heads give the mean share of shared gates. (Manuscript Table 15.)

ModelShared gates 22.9%Shared gates 40.2%Shared gates 53.4%
ShareGate (ours)98.79→99.7490.17→99.4489.62→98.96
ABGNN95.46→96.6186.58→94.8882.19→90.81
DAGNN-RE87.46→88.1872.92→83.9658.48→71.26
GAT90.49→93.3181.75→91.5376.70→84.82
GCN77.60→89.5666.90→84.3455.37→71.88
GraphSAGE85.24→88.2671.96→84.1758.61→71.65
HGCN66.76→95.8974.50→91.9672.99→85.55
NeuroMatch67.66→87.4539.08→78.9128.36→62.08
ReVEAL59.48→89.2945.51→79.2339.76→62.70
WideGate92.03→96.6287.81→95.4984.52→91.95

Table R2. Certification by VDR. (Manuscript Table 16.)

SettingADD cores certifiedSUB cores certified
No sharing (L0)21/2121/21
~0.23 (22.9%)21/2121/21
~0.46 (40.2%)21/2121/21
~0.73 (53.4%)21/2121/21

[W1, Q2] Can you run a sweep that varies sharing rather than dilution?

Following the reviewer's suggestion, we vary the number of operations that reuse the same intermediate expression. Actual shared gates rise from 22.9% to 40.2% and 53.4% (mean over the 21 netlists), and ShareGate stays the most accurate on shared gates and certifies every adder and subtractor core at every level (Table R1,R2).

The settings use three arithmetic identities :

1.the comparator reuses the borrow chain of the subtractor;

2.the comparator is derived entirely from the subtractor;

3.the subtractor reuses the adder through a − b = (a + b) − 2b.

This experiment keeps the host logic and removes the multiplier. The multiplier holds about three quarters of the gates and shares only its inputs with the other cores, so with it present actual shared gates stay below 12.4% on average. The abstract and Section 4 now report both experiments.

AuthorsOpenReview ↗

Response to Reviewer nvii (2/3) - Q3

[W2, Q3] What does the baseline toolbox do, and how long do the baselines take with the same refinement?

The toolbox works in two stages. For every case it first rebuilds the label field and the structural state of the whole netlist, which takes a median of 123–140 s whatever the label quality. It then searches the label field for a closed region and calls CEC on each attempt, which takes a median of 13–18 s for seven baselines, 62.0 s for ReVEAL and 173.3 s for ABGNN. The first stage explains the 130–150 s plateau. Per-baseline stage times are in manuscript Table 17.

With the same proposal-and-check refinement, the nine baselines reach a certified result in 4.7–5.6 s at ρ = 0 and 27.8–28.9 s at ρ = 0.75, against 3.9 s and 21.7 s for ShareGate (Table R5). The 6.0–34.5× margin of Table 3 therefore measures VDR against the toolbox, as the reviewer pointed out. We have revised the abstract and Section 4.6 to say so. ShareGate is the fastest at three of the four levels; at ρ = 0.5, GCN + VDR is slightly faster (18.81 s against 19.05 s).

Table R3. Stage times of the original toolbox (timing panel of the paper: 5 netlists × 4 ρ, recorded per case). (Manuscript Table 17.)

BaselinePer-case state preparation, median (s)Closed-region search + CEC, median (s)
ABGNN140.4173.3
DAGNN-RE129.917.8
GAT126.217.4
GCN124.417.4
GraphSAGE131.217.0
HGCN127.817.7
NeuroMatch131.018.0
ReVEAL126.662.0
WideGate123.212.8

Table R4. Original toolbox: median time (s) to a certified result (manuscript Table 3).

Methodρ = 0ρ = 0.25ρ = 0.5ρ = 0.75
ShareGate (ours)3.9310.0919.0521.65
ABGNN189.8160.8167.7163.7
DAGNN-RE146.6151.8145.0142.9
GAT138.3149.3148.9140.5
GCN141.5148.1138.2142.0
GraphSAGE145.2153.9145.8146.7
HGCN146.1140.9144.9151.7
NeuroMatch146.8147.9152.6144.9
ReVEAL137.1196.0187.1178.6
WideGate135.6134.2133.9129.7

Table R5. Same proposal-and-check refinement: stage times (s). Baselines: graph preparation + forward pass, VDR input preparation, VDR engine. ShareGate: forward, engine, verify (manuscript Table 13). (Manuscript Table 18.)

Methodρ = 0ρ = 0.25ρ = 0.5ρ = 0.75
ShareGate (ours)0.380 + 0.609 + 2.9420.631 + 6.335 + 3.1250.825 + 12.657 + 5.5691.028 + 16.560 + 4.062
ABGNN0.04 + 0.96 + 3.760.05 + 1.20 + 14.830.07 + 1.58 + 20.440.08 + 1.57 + 26.95
DAGNN-RE0.05 + 0.97 + 3.850.07 + 1.19 + 14.950.09 + 1.52 + 19.150.12 + 1.78 + 26.76
GAT0.04 + 0.95 + 3.700.05 + 1.19 + 14.840.07 + 1.52 + 20.570.08 + 1.73 + 26.66
GCN0.04 + 0.95 + 3.900.05 + 0.66 + 13.620.06 + 1.48 + 17.270.09 + 1.80 + 26.22
GraphSAGE0.04 + 0.95 + 3.900.05 + 1.19 + 15.140.07 + 1.53 + 20.580.08 + 1.77 + 26.76
HGCN0.08 + 0.95 + 3.780.10 + 1.20 + 14.940.13 + 1.52 + 19.620.15 + 1.80 + 26.87
NeuroMatch0.03 + 0.94 + 3.860.05 + 1.18 + 14.920.06 + 1.52 + 20.810.08 + 1.78 + 26.47
ReVEAL0.83 + 0.95 + 3.781.11 + 1.18 + 14.791.45 + 1.43 + 19.611.77 + 1.78 + 24.25
WideGate0.04 + 0.95 + 3.700.05 + 1.19 + 8.950.07 + 1.53 + 20.690.08 + 1.76 + 27.04
AuthorsOpenReview ↗

Response to Reviewer nvii (3/3) - Q4

Q4.[W2, Q4] Does refinement work on the ABGNN and DAGNN-RE fields, and is the classifier then irrelevant?

We applied VDR to the ABGNN and DAGNN-RE fields and the other seven baseline fields (Tables R6–R9). Under identical refinement, VDR raises Acc, F1, and MCC for every baseline at every ρ , and BAcc in 31 of 36 cells; the five exceptions lie at ρ ≤ 0.25 and stay within 1.01 points. For example, at ρ = 0, ABGNN rises from 95.89% to 96.14% and DAGNN-RE from 94.52% to 96.48%. Thus, the main contribution is verification-driven refinement, consistent with Section 3.1.

The classifier is not irrelevant. VDR rewrites only certified regions, while all other gates retain the classifier's predictions. Furthermore, VDR’s candidate regions are derived not only from the multi-bit output cone but also rely on "cosine density clustering in the learned embedding space" (Section 3.3, C.2); thus, the final accuracy remains dependent on the prior. With this consistent refinement, the complete ShareGate system ranks first across all sixteen cells in Table 1: at ρ = 0, it achieves 97.02% accuracy (compared to 96.66% for the best refined baseline, GraphSAGE + VDR), and at ρ = 0.75, it reaches 92.34% accuracy and 80.82 MCC (compared to 91.15% and 78.49%, respectively). The 83.8% figure cited by the reviewer refers to ShareGate’s raw classifier at ρ = 0. The outputs generated by ShareGate are the optimized results; the prior information serves to adapt to VDR and provides a foundational basis at every level.

Table R6. Observation gates, ρ = 0 . (Manuscript Table 19.)

PriorAccF1MCCbAcc
ABGNN + VDR96.1490.6391.8592.39
DAGNN-RE + VDR96.4890.7992.4792.67
GAT + VDR95.8089.0690.8791.14
GCN + VDR95.7887.9990.5089.93
GraphSAGE + VDR96.6691.0592.8993.05
HGCN + VDR95.1185.4588.8288.21
NeuroMatch + VDR90.7779.3879.0281.24
ReVEAL + VDR89.7570.0675.1572.53
WideGate + VDR95.3587.7789.8690.51

Table R7. Observation gates, ρ = 0.25 . (Manuscript Table 19.)

PriorAccF1MCCbAcc
ABGNN + VDR89.9168.1074.8572.32
DAGNN-RE + VDR90.0164.5576.3372.74
GAT + VDR89.7163.0175.7272.50
GCN + VDR90.7763.8877.7072.77
GraphSAGE + VDR90.9664.7078.2973.31
HGCN + VDR91.2469.7778.3374.59
NeuroMatch + VDR84.9258.7963.6665.28
ReVEAL + VDR87.5359.5770.7667.49
WideGate + VDR89.4964.0174.7371.67

Table R8. Observation gates, ρ = 0.5 . (Manuscript Table 19.)

PriorAccF1MCCbAcc
ABGNN + VDR88.3265.7570.0968.65
DAGNN-RE + VDR87.3255.9369.8266.93
GAT + VDR89.6162.6874.5871.31
GCN + VDR90.0059.5075.8971.06
GraphSAGE + VDR90.4859.7676.9971.06
HGCN + VDR88.9565.5473.0971.26
NeuroMatch + VDR84.3656.1961.7063.70
ReVEAL + VDR87.2960.2169.5767.10
WideGate + VDR88.4362.5571.1869.52

Table R9. Observation gates, ρ = 0.75 . (Manuscript Table 19.)

PriorAccF1MCCbAcc
ABGNN + VDR87.4165.5467.7567.30
DAGNN-RE + VDR85.9055.4865.8364.79
GAT + VDR90.0165.4075.0072.23
GCN + VDR90.2260.4876.1471.52
GraphSAGE + VDR91.1561.3878.4971.82
HGCN + VDR88.6965.9571.9270.25
NeuroMatch + VDR84.3856.6561.4063.54
ReVEAL + VDR87.2062.2868.7367.02
WideGate + VDR88.4965.2470.7269.36

Comments on the paper

1 comment

Dear Reviewers,

The authors have submitted their rebuttal. Please review it and share your comments. If anything is unclear or requires clarification, please let the authors know as soon as possible. The author–reviewer discussion period ends on 18 November.

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.