acceptodds
Under review as a conference paper at ICLR 2027

EvoPCB: Evolving PCB Physical Design to Expert Alignment via Policy Optimization

Abstract

Despite substantial progress in automated printed circuit board (PCB) physical design, mainly including placement and routing (PnR), industrial physical design still relies heavily on human expert engineers. Existing frameworks can generate design-rule-compliant designs, but conventional objectives alone do not ensure satisfactory performance or compliance with industrial standards. Moreover, existing frameworks often cannot adapt to user specifications or domain-specific constraints, which also limits their applicability. To handle these problems, we propose EvoPCB, an end-to-end, LLM-based PCB physical design framework. EvoPCB introduces RePCB, a performance-aware supervised representation model that embeds PnR topology and measures similarity to expert layouts. A two-stage specialization, supervised fine-tuning (SFT) followed by customized decoupled clip and dynamic sampling policy optimization (DAPO), assists EvoPCB in generating high-quality initial designs. EvoPCB then refines the layouts through a multi-agent and multi-objective script evolution method, MM-Evolve, which coordinates geometric and multi-physics optimization. Experimental results demonstrate that EvoPCB achieves the best overall performance compared to existing state-of-the-art methods across all evaluated metrics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
LLM, Physical Design, PCB, Policy Optimization, Evolving

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_WZ21OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents EvoPCB, an LLM-based framework for end-to-end PCB physical design that formulates placement and routing as script generation and optimization over domain-specific APIs. The framework comprises three main components. First, RePCB jointly encodes component connectivity, placement geometry, and routing images, with supervision derived from wirelength and via count. Second, Bayesian optimization constructs layout–script training pairs, which support SFT and a customized DAPO procedure for generating initial designs. Finally, MM-Evolve iteratively refines the designs through topology, thermal, and electromagnetic agents using simulation feedback.

Critical strengths

The problem is practically relevant, and the API-based formulation makes sense for the task PCB physical design involves discrete decisions, geometric constraints, and competing physical objectives. Having the LLM generate calls to existing placement and routing tools gives it a concrete role in the design process: selecting and configuring tools, while the tools carry out the geometric operations.

The data construction addresses a concrete limitation of the available supervision. Expert PCB layouts often lack the scripts that produced them. Searching over a constrained API space to recover executable scripts is a practical way to turn these layouts into training examples. RePCB’s use of connectivity graphs and routing images is also a reasonable choice for representing the different aspects of a PCB layout.

Critical weaknesses

RePCB’s supervision remains largely determined by conventional metrics. I am not yet convinced that its similarity score captures expert design preferences beyond wirelength and via count. In Equation (4), both the regression targets and the pairwise similarity targets are derived from these two quantities. Two layouts with identical wirelength and via count therefore receive a target similarity of one, even if their routing structures differ substantially. The EM and thermal prediction results in Table 7 are useful, but downstream prediction accuracy does not directly establish that the similarity score reflects expert preferences.

The benefit of the DAPO customization has not been established. There is no controlled comparison with standard DAPO using the same PCB reward components and training budget. In particular, the contributions of the LBI filter and the modified advantage remain unclear. LBI checks whether a layout is legal and differs from the reference; this condition alone does not ensure variation in rewards within a group. The paper should clarify the actual filtering procedure and provide evidence that these changes improve training efficiency or final design quality.

Questions for authors

How does performance change if RePCB similarity is replaced by a similarity score computed directly from the normalized wirelength and via-count features used in Equation (4), while keeping the rest of the pipeline unchanged?

Can RePCB distinguish layouts with similar wirelength and via count but different topology or expert-assessed quality? A controlled comparison of such layouts would help clarify what the representation captures beyond its supervised metrics.

How does the customized DAPO compare with standard DAPO when using the same initial checkpoint, training data, PCB reward components, and rollout budget? It would also be helpful to separate the effects of LBI and the modified advantage.

10 comments

We would like to express our gratitude for your thoughtful review and recognition of our work. Your comments have guided us in clarifying our methodology. We have added more detailed evidence to the anonymous repo: https://anonymous.4open.science/r/EvoPCB-ICLR-2027-8E0E/Rebuttal_EvoPCB.md to address the reviewers' concerns. Key additional content will be added to the paper.

1. RePCB Similarity Analysis

In the reward ablation below, we keep the group-normalized WL/via reward riWV=α(wˉ−wi)/(σw+δ)+β(vˉ−vi)/(σv+δ)r_i^{\mathrm{WV}}=\alpha(\bar w-w_i)/(\sigma_w+\delta)+\beta(\bar v-v_i)/(\sigma_v+\delta) fixed and change only the similarity term. The four configurations use: (i) no similarity term; (ii) the analytic WL/via similarity target pip_i from Eq. (4); (iii) centered RePCB similarity si−sˉs_i-\bar s; and (iv) uncentered RePCB similarity sis_i. Their complete rewards are therefore riWVr_i^{\mathrm{WV}}, riWV+γpir_i^{\mathrm{WV}}+\gamma p_i, riWV+γ(si−sˉ)r_i^{\mathrm{WV}}+\gamma(s_i-\bar s), and riWV+γsir_i^{\mathrm{WV}}+\gamma s_i, respectively. The last two configurations have the same within-group similarity ordering and differ by the group-wide offset γsˉ\gamma\bar s, which can affect finite-sample, filtered, and clipped policy updates. The remainder of the pipeline is unchanged.

ConfigurationWL(mm) ↓\downarrowVia ↓\downarrowSEM↓\mathcal{S}_{\mathrm{EM}}\downarrowSTh↓\mathcal{S}_{\mathrm{Th}}\downarrowRePCB similarity to Expert Designs ↑\uparrow
WL+Via only (no similarity term)1176.950.37.062.5840.541
WL+Via + analytic metric similarity (Eq. 4 target)1191.348.94.872.2140.603
WL+Via + centered RePCB similarity, si−sˉs_i-\bar{s}1217.149.23.542.1370.821
WL+Via + uncentered RePCB similarity, sis_i (ours)1208.547.12.781.7360.873

Direct metric rewards produce slightly shorter WL, but all three similarity-based configurations improve EM/thermal costs over WL+Via only. Centered RePCB further improves RePCB similarity and EM cost over analytic metric similarity, while the uncentered variant gives the best via count, EM/thermal costs, and RePCB similarity. These results support the empirical utility of the learned RePCB signal and of the uncentered design in the reported pipeline; they do not establish that RePCB directly optimizes EM/thermal objectives. The RePCB-similarity column measures agreement under the learned representation and is not an independent quality measure.

2. Discrimination Ability of RePCB

The topology-sensitivity study below evaluates 30 layout pairs per category: one group with clearly different WL/via counts but similar PnR topology, and another group with similar WL/via counts but different PnR topology. Using the same fixed 0-1 rubric for all shuffled pairs, topology similarity was rated independently by three senior PCB engineers (scores averaged) and by GPT-6 Astra, from route geometry, relative component placement, and layer-transition patterns. RePCB similarity is the normalized cosine similarity between the corresponding embeddings.

Type∣ΔWL∣\vert\Delta\mathrm{WL}\vert(%)∣ΔVia∣\vert\Delta\mathrm{Via}\vert(%)Expert Topology Similarity ↑\uparrowGPT-6 Astra Topology Similarity ↑\uparrowRePCB Similarity ↑\uparrow
Different WL/Via, Similar Topology38.5±13.126.4±8.80.91±0.040.93±0.020.904±0.031
Similar WL/Via, Different Topology1.7±0.91.9±0.70.26±0.060.23±0.040.312±0.067

RePCB similarity remains high when topology is similar despite large WL/via differences, and decreases when topology differs despite similar WL/via values, in agreement with the independent engineers' ratings. These results provide evidence that RePCB is sensitive to topology beyond its two scalar supervision targets on the evaluated pairs.

3. Effectiveness of Customized DAPO

All configurations use the same initial checkpoint, training data, PCB reward components, and evaluation protocol. “Naive DAPO” is the standard DAPO comparison; “Customized DAPO w/o LBI” changes the advantage; and the final row adds LBI. This sequential comparison evaluates the advantage change followed by LBI.

RL ConfigurationWL(mm) ↓\downarrowVia ↓\downarrowSEM↓\mathcal{S}_{\mathrm{EM}}\downarrowSTh↓\mathcal{S}_{\mathrm{Th}}\downarrowLegal script(%) ↑\uparrow
w/o DAPO1389.661.46.212.24788.7
Naive DAPO1357.958.85.652.08689.8
Customized DAPO w/o LBI1291.351.03.411.86489.2
Customized DAPO (Ours)1208.547.12.781.73699.1

Here, a legal script is one that executes successfully and produces a DRC/LVS-valid layout. The final model's 99.1% rate corresponds to 991 legal outputs among 1,000 held-out PCB design tasks, with one generated initial script per task. LBI retains a rollout in the loss only when it is legal and differs from the reference under the implementation's coarse metric-level WL/via criterion, thereby focusing the finite update budget on non-equivalent outcomes under the two primary scalar routing objectives. Metric matching does not imply script, layout, or topology equality, and matched rollouts may still contain structural information; we therefore regard this filter as an empirical design choice supported by the ablation rather than a theoretically necessary criterion. LBI also does not itself guarantee reward variation. The comparison from Naive DAPO to Customized DAPO w/o LBI measures the modified advantage, and the final comparison isolates LBI. LBI raises the empirical legal-script rate from 89.2% to 99.1%, although illegal scripts are filtered rather than assigned an explicit negative advantage.

Conclusion

Thank you again for your valuable and constructive feedback. Your comments have substantially improved the clarity of our paper. We would be grateful if you could reevaluate our work and reconsider the score in light of these clarifications and additional results.

Dear Reviewer,

We hope our clarification has addressed your concerns. We have added further experiments to the anonymous repository: https://anonymous.4open.science/r/EvoPCB-ICLR-2027-8E0E/Rebuttal_EvoPCB.md. We welcome any feedback you may have, and we would be very grateful if you could reconsider your score in light of our clarification.

Best regards,

All authors

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

The paper proposes EvoPCB, an LLM-based framework for PCB placement and routing. Instead of placing components directly, the LLM writes a script that calls existing placement and routing algorithms, such as simulated annealing, MILP, differentiable placement and A* routing. EvoPCB has three stages. First, Qwen3-14B is fine-tuned on example scripts. Expert layouts come without scripts, so the authors use Bayesian optimization to find scripts whose layouts are close to the expert layouts. Second, the model is trained with reinforcement learning, using a customized version of DAPO. The reward favors scripts whose layouts have shorter wirelength, fewer vias, and higher similarity to the expert layout. The similarity is computed by RePCB, a GraphSAGE and CNN model that embeds a layout and is trained to predict its wirelength and via count. Third, a multi-agent stage (MM-Evolve) refines the script. Agents built on Qwen3-VL-8B inspect layout images and the results of electromagnetic (openEMS) and thermal (OpenFOAM) simulations. A coordinator agent then decides which of their suggested edits to apply. The EM cost combines insertion loss and crosstalk, and the thermal cost combines peak temperature and peak temperature gradient. On 11 boards with 35 to 97 components, the authors report the best average wirelength, via count, EM cost, and thermal cost against PCB-PR-App, NS-Place, PCBAgent, DeepPCB, and GPT-6 Astra.

Critical strengths
  1. Having the LLM write scripts that call proven solvers, instead of coordinates, is a sensible way to use LLMs for PCB design.
  2. The pipeline covers placement, routing, design rule checks and multi-physics evaluation, and the comparison includes a commercial tool.
  3. The appendices describe the APIs and the BO-based data preparation in detail.
Critical weaknesses
  1. The EM and thermal comparison is not fair. MM-Evolve optimizes the same openEMS and OpenFOAM costs that are used for evaluation, while no baseline sees these simulators. Lower EM and thermal costs are therefore expected. A fair test would give a baseline the same simulation-in-the-loop refinement, or run a simple search over script parameters with the same simulation budget.
  2. The abstract claims the best performance "across all evaluated metrics", and the conclusion says EvoPCB "outperforms all existing state-of-the-art baselines". This holds only for the averages. All results come from one seed (2026), and no repeated runs are reported, so run-to-run variation is unknown.
  3. The RL objective has design problems. In Eq. (8), the wirelength and via terms are centered within the group, but the similarity term γ·sim is not. This adds a positive offset to every legal script, biasing the update toward increasing the likelihood of all legal scripts in the group. The LBI indicator removes illegal scripts from the update instead of giving them a negative advantage, so the policy gets no direct signal to avoid illegal outputs. Please justify these choices or compare them with standard DAPO.
  4. Eq. (1) defines a weighted cost with weights μ1 to μ4, but the weights are never given, and it is unclear where this cost is used.
Questions for authors
  1. How many scripts are sampled per test board, and how is the reported one selected?
  2. What are the runtime and the number of simulator calls of MM-Evolve per board?
  3. Where do the expert layouts come from, and how was overlap with the test boards checked?
  4. How are the critical nets, the operating frequency and the component power defined for each board?
9 comments

We are grateful for the evaluation and constructive critiques. Incorporating your suggestions has allowed us to strengthen key methodological aspects. A structured response to each point follows. We have added more detailed evidence to the anonymous repo: https://anonymous.4open.science/r/EvoPCB-ICLR-2027-8E0E/Rebuttal_EvoPCB.md to address the reviewers' concerns. Key additional content will be added to the paper.

1. Utilization of Simulation

We agree with the fairness concern. We therefore add simulator-assisted baselines below. The average times for one rapid EM simulation and one thermal simulation were 131.5s and 13.8s, respectively. MM-Evolve performs 3.7 full EM and 2.4 full thermal simulations per board on average, yielding an end-to-end runtime of 1045.89s.

MethodTime(s) ↓\downarrowRatio
PCB-PR-App344.90.33×
NS-Place+FreeRouting8682.78.30×
PCBAgent+FreeRouting1788.21.71×
DeepPCB7378.37.05×
GPT-6 Astra2620.02.51×
EvoPCB1045.891.00×

Among the existing baselines, only PCBAgent can be adapted to search over scripts using simulation feedback. We also add a Bayesian optimization (BO) search over the same script-template space. Both simulator-assisted searches remain worse than EvoPCB, while the comparison with EvoPCB w/o MM-Evolve quantifies the gain and added runtime of direct physical refinement.

ImplementationWL(mm) ↓\downarrowVia ↓\downarrowSEM↓\mathcal{S}_{\mathrm{EM}}\downarrowSTh↓\mathcal{S}_{\mathrm{Th}}\downarrowEM calls ↓\downarrowThermal calls ↓\downarrowTime(s) ↓\downarrow
PCBAgent+Simulation1791.1156.79.282.07756.356.375643.0
BO+Simulation4271.9294.212.264.21870.070.012020.2
EvoPCB w/o MM-Evolve1219.252.28.742.57800301.8
EvoPCB1208.547.12.781.7363.72.41045.89

The results show that providing simulator access to these searches does not by itself reproduce EvoPCB's quality. EvoPCB has better values on all four quality metrics and shorter runtime under the reported configurations. Even without MM-Evolve, EvoPCB improves on BO + Simulation on all four metrics and on PCBAgent + Simulation on WL, via count, and EM cost; this supports the value of the specialized initialization.

These are comparisons with simulator access. PCBAgent+Simulation averages 56.3 EM and 56.3 thermal simulations per board, while BO+Simulation uses 70 EM and 70 thermal simulations per board. For each baseline, the EM and thermal call counts are equal because both simulators are invoked in every search iteration. In comparison, EvoPCB averages 3.7 EM and 2.4 thermal simulations per board during MM-Evolve. Thus, the simulator-assisted baselines use substantially larger physical-simulation budgets under the reported configurations but remain worse on the reported quality metrics.

BoardEM callsThermal callsMM-Evolve time (s)
P153996.4
P232599.7
P332601.2
P453995.1
P543805.6
P622408.9
P742790.3
P832598.8
P932600.5
P1053997.0
P1142791.6
Average3.72.4744.1

2. Stability across Random Seeds

Seed 2026 is the default setting used for the original table. To assess run-to-run variability separately, we additionally evaluated 30 random seeds and report the mean, standard deviation, and range as follows:

StatisticWL(mm) ↓\downarrowVia ↓\downarrowSEM↓\mathcal{S}_{\mathrm{EM}}\downarrowSTh↓\mathcal{S}_{\mathrm{Th}}\downarrowTime(s) ↓\downarrow
EvoPCB (default)1208.547.12.781.7361045.89
mean±std1214.8±18.747.6±1.82.91±0.421.748±0.0281048.7±22.9
min/max1187.9/1248.645.1/50.82.24/3.631.704/1.7971012.4/1088.3

We keep the benchmark boards, objective weights, iteration limits, and simulator settings fixed, and vary only the seeds of the stochastic EDA operations, namely simulated-annealing placement and rip-up-and-reroute. Different seeds can therefore produce different placement trajectories and routing solutions, which may lead MM-Evolve along different refinement paths and cause variation in WL, via count, EM/thermal costs, and runtime. The EM and thermal solvers themselves are deterministic for a fixed layout and configuration, so their variation is induced indirectly by the resulting layouts.

The 30-seed results further show that even the largest observed EvoPCB averages for WL (1248.6), EM (3.63), and thermal cost (1.797) are below the corresponding best reported baseline averages (1346.1, 7.30, and 2.001). These observations support robustness to the evaluated EvoPCB seeds, without accounting for baseline run-to-run variability.

Meanwhile, we will revise the abstract and conclusion to claim the best average performance among the compared methods on the evaluated benchmark, rather than universal superiority over all existing methods.

3. Implementation of Customized DAPO

Our advantage uses centered WL/via terms and an uncentered RePCB similarity term. The latter contains both relative similarity differences and a group-mean offset. Centering it would retain reference-based comparisons while removing that offset. With filtering, finite sampling, and clipping, this can change the update. We therefore treat the uncentered term as an empirical design choice, rather than claiming that an offset is necessary for reference alignment. The DAPO ablation below evaluates standard DAPO, the modified advantage without LBI, and the combined customization; it does not separately test similarity centering.

EvoPCB already makes illegal outputs rare through two complementary safeguards. First, the model generates scripts from a reliable PnR template composed of typed APIs with validated parameters and runtime safety constraints, which substantially restricts malformed or unsafe actions. Second, Legal-But-Imperfect (LBI) serves as an execution-time eligibility gate during DAPO: every generated script is executed and checked before it can contribute to training, and a rollout is retained only if execution succeeds, the resulting layout is valid, routed wirelength is positive, no shorts or clearance violations are detected, and placement overlap and out-of-bounds area do not exceed the reference tolerances. Thus, the API layer reduces the probability of producing an illegal layout, while LBI prevents any residual invalid rollout from entering the group statistics or policy update. We do not claim that syntax and API constraints alone mathematically guarantee legality.

Given these safeguards and the observed 99.1% legal-script rate (991 successful designs among 1,000 held-out PCB design tasks), an additional negative reward would target a very small residual set while introducing an arbitrary penalty scale that could interfere with the WL, via, and topology objectives. We therefore use the LBI hard gate instead of adding a separately tuned legality penalty. Under LBI, illegal or unparsable scripts receive zero reward and are excluded from the group statistics and policy update; group means and standard deviations are computed only from legal, non-reference-matching rollouts. If a group has no eligible rollout, it contributes no update. The released implementation's reference-match test is a coarse metric-level eligibility heuristic based on wirelength and via count. It focuses the finite update budget on legal rollouts with outcomes that differ from the reference under these two primary scalar routing objectives. Importantly, matching these metrics does not imply script, layout, or topology equality, and matched rollouts may still contain structural information. We therefore treat this filter as an empirical design choice supported by the LBI ablation, rather than as a theoretically necessary criterion. We will revise the paper to state these implementation details explicitly.

All variants below use the same initial checkpoint, training data, PCB reward components, and evaluation protocol. “Naive DAPO” is the standard DAPO baseline; “Customized DAPO w/o LBI” changes the advantage; and the final row adds LBI. These sequential comparisons support the observed gains from the modified advantage and from adding LBI. A legal script is one that executes successfully and produces a DRC/LVS-valid layout.

RL ConfigurationWL(mm) ↓\downarrowVia ↓\downarrowSEM↓\mathcal{S}_{\mathrm{EM}}\downarrowSTh↓\mathcal{S}_{\mathrm{Th}}\downarrowLegal script(%) ↑\uparrow
w/o DAPO1389.661.46.212.24788.7
Naive DAPO1357.958.85.652.08689.8
Customized DAPO w/o LBI1291.351.03.411.86489.2
Customized DAPO (Ours)1208.547.12.781.73699.1

4. Problem Formulation Explanation

Moreover, the weights in Eq. (1) provide an abstract formulation of the multi-objective problem rather than fixed inputs that users must specify. Eq. (1) expresses the overall multi-objective goal, whereas BO data preparation and DAPO use the distinct objectives specified in their respective sections. The four weights in Eq. (1) cannot be inferred from the three BO/DAPO weights. We agree that this was unclear and will clarify it in the paper.

5. More Implementation Details

  1. Scripts and Selection. After DAPO, we conducted a separate held-out evaluation of the specialized model on 1,000 PCB design tasks and obtained 991 legal outputs, corresponding to a 99.1% legal-script rate. Each task generated one initial PnR script, so the denominator is the full set of 1,000 generated scripts, with no best-of-NN selection. This should not be confused with MM-Evolve: during refinement, the coordinator forms a beam of K=6K=6 conflict-free candidate scripts. The reported benchmark result is the final iteratively refined design, not the best of multiple independently sampled initial scripts.
  2. MM-Evolve cost. During MM-Evolve refinement, the coordinator generates up to K=6K=6 conflict-free candidate scripts and screens them with fast rule-based checks (DRC/LVS feasibility, WL, via count, and RePCB similarity) that require no full simulation; the single most promising candidate is then selected for a full EM simulation. MM-Evolve performs 3.7 full EM and 2.4 full thermal simulations per board on average, with at most five full EM simulations per board.
BoardEM callsThermal callsMM-Evolve time (s)
P153996.4
P232599.7
P332601.2
P453995.1
P543805.6
P622408.9
P742790.3
P832598.8
P932600.5
P1053997.0
P1142791.6
Average3.72.4744.1

Runtime is defined as end-to-end wall-clock design time per board, averaged over the 11 benchmarks and including physical refinement. The measured runtime comprises approximately 301.8s for the initial stage and 744.1s for refinement, consistent with the 1045.89s total. The individual operation latencies are:

OperationAverage time per call (s)↓\downarrow
EM simulation131.5
Thermal simulation13.8
Multimodal analysis31.2
PnR script generation20.5
  1. The expert layouts come from a top-tier company in the industry. The training and test sets are disjoint. The tables below present the dataset distribution according to the number of components, nets, and layers.
  2. In industrial applications, nets corresponding to critical signals are explicitly marked. Operating frequencies and component power are derived from the libraries associated with the dataset and held fixed across methods.
Subset≤40 comps41–80 comps>80 comps≤50 nets51–100 nets>100 nets2 layers4–6 layers>6 layers
SFT Dataset615129842813059201161492665184
DAPO Dataset1392317611733025238233435361650555

Conclusion

Thank you for your thoughtful feedback. Your comments have helped us improve the paper. We have provided detailed supplementary information in response to your questions. We believe these changes strengthen the paper and would appreciate your updated score.

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

EvoPCB addresses an LLM-based framework for PCB placement and routing. It formulates physical design as executable script generation, allowing an LLM to invoke existing placement and routing algorithms through predefined APIs. RePCB combines GraphSAGE-based placement encoding and CNN-based routing encoding to learn performance-aware representations. Supervised fine-tuning followed by customized DAPO produces initial design scripts, while MM-Evolve uses multiple specialized agents to refine layouts with geometric, electromagnetic, and thermal feedback. Experiments on 11 PCB designs report improvements in wirelength, via count, EM cost, and thermal cost, with ablations examining individual components.

Critical strengths

1.The script-generation formulation provides a practical interface between LLMs and PCB design tools. Rather than directly predicting component coordinates and routing geometries, the model controls algorithm selection, API invocation sequences, and parameter configurations through executable scripts. This formulation makes effective use of existing EDA algorithms and provides a meaningful setting for combining supervised specialization with policy optimization. 2.The framework jointly considers geometric and multi-physics objectives. EvoPCB extends optimization beyond wirelength and via count to include electromagnetic and thermal performance. Its multi-agent refinement process offers an interesting approach to coordinating potentially conflicting design objectives. The reported results and ablations suggest that MM-Evolve contributes substantially to multi-physics improvements, making this integration a consequential part of the paper’s practical contribution.

Critical weaknesses

1.The central expert-alignment claim is not sufficiently distinguished from conventional metric optimization. RePCB is supervised using wirelength and via count, and its pairwise similarity target is constructed from differences in those metrics. It is therefore unclear whether the learned similarity captures expert-specific preferences beyond geometric performance, such as functional grouping or critical-signal placement. Accurate prediction of downstream EM or thermal metrics does not by itself establish agreement with human expert preferences. Moreover, the SFT reference scripts are obtained through Bayesian optimization to reproduce expert layouts, rather than recording expert-generated design sequences. Learning to reproduce reference outcomes should therefore be distinguished from learning expert design strategies. A comparison against simpler rewards based on normalized wirelength and via-count differences would help establish RePCB’s additional value. Independent expert preference comparisons would further support the stronger alignment claim, or the claim could be narrowed to the demonstrated form of performance-based alignment. 2.The evidence does not yet adequately establish the contribution, computational efficiency, and generalization of the final multi-physics improvements. DAPO uses wirelength, via count, and RePCB similarity, whereas the final objectives additionally include EM and thermal performance. Since RePCB supervision is primarily geometric, the relationship between these surrogate rewards and downstream physical quality remains unclear. The substantial EM and thermal degradation without MM-Evolve supports the importance of that component, but leaves uncertainty about how much earlier stages contribute beyond providing a geometrically optimized initialization. This matters because iterative simulation-guided refinement may itself account for a substantial part of the improvement. Without simulation counts, latency, total runtime, and comparisons under matched budgets, the efficiency and fairness of the reported gains are difficult to assess. The scale distributions of the 2,341 SFT and 5,741 DAPO expert-layout datasets are not reported, and the experiments use a fixed random seed. These limitations leave the claimed breadth and stability of the improvements insufficiently characterized. Larger public benchmarks offer possible extensions, although adapting them to the end-to-end task would require additional work.

Questions for authors

1.Expert alignment: What evidence shows that RePCB captures expert preferences beyond wirelength and via count? How does its reward compare with a simpler similarity reward based directly on normalized geometric metrics? Is there evidence of agreement with independent expert preferences? 2.Surrogate rewards and physical performance: What is the relationship between RePCB similarity and downstream EM/thermal quality? Under a controlled MM-Evolve budget, how much do the RePCB and DAPO stages improve the final result compared with simpler initializations? 3.Computational efficiency and comparison budgets: What are the EM and thermal simulation latencies, evaluation counts per design, and total end-to-end runtimes? Do the reported advantages persist under matched wall-clock time or physical-evaluation budgets? 4.Generalization and robustness: Please report the component-count, net-count, and layer-count distributions of the SFT and DAPO datasets. Can the authors provide representative multi-seed results and evidence on larger designs to clarify the stability and scope of generalization?

6 comments

Thank you for your constructive comments and recognition of the core contributions of our work. We greatly appreciate your thoughtful feedback, which has helped us further refine and clarify several key aspects of our paper. Our anonymous repo: https://anonymous.4open.science/r/EvoPCB-ICLR-2027-8E0E/Rebuttal_EvoPCB.md reports more detailed experimental results. We will reproduce the relevant results below and will incorporate them, together with the corresponding scope clarifications, into the paper.

1. The Significance of Expert Alignment

We agree that outcome alignment should be distinguished from recovering an expert's design process. RePCB is supervised by WL and via count, and the BO-derived SFT scripts reproduce expert layouts rather than record engineers' actions. We therefore do not claim to recover expert policies or the full range of human preferences; our claim is performance-driven alignment with expert reference layouts, with topology sensitivity assessed separately below. We will revise the paper consistently to distinguish reference-outcome alignment from learning expert strategies or preferences. To assess RePCB's added value, the reward ablation keeps the group-normalized WL/via terms fixed and changes only the similarity term. Let riWV=α(wˉ−wi)/(σw+δ)+β(vˉ−vi)/(σv+δ)r_i^{\mathrm{WV}}=\alpha(\bar w-w_i)/(\sigma_w+\delta)+\beta(\bar v-v_i)/(\sigma_v+\delta), pip_i be the analytic WL/via similarity target from Eq. (4), and sis_i be RePCB similarity to the paired reference. The configurations respectively use riWVr_i^{\mathrm{WV}} only, riWV+γpir_i^{\mathrm{WV}}+\gamma p_i, riWV+γ(si−sˉ)r_i^{\mathrm{WV}}+\gamma(s_i-\bar s), and riWV+γsir_i^{\mathrm{WV}}+\gamma s_i (ours). The centered and uncentered RePCB terms induce the same within-group similarity ordering, but the uncentered form additionally contributes the group-wide offset γsˉ\gamma\bar s. This offset can affect finite-sample, filtered, and clipped policy updates; we therefore evaluate it as an empirical design choice rather than claiming that it supplies additional rollout-level ranking information.

ConfigurationWL(mm) ↓\downarrowVia ↓\downarrowSEM↓\mathcal{S}_{\mathrm{EM}}\downarrowSTh↓\mathcal{S}_{\mathrm{Th}}\downarrowRePCB similarity to Expert Designs ↑\uparrow
WL+Via only1176.950.37.062.5840.541
WL+Via + analytic metric similarity1191.348.94.872.2140.603
WL+Via + centered RePCB similarity1217.149.23.542.1370.821
WL+Via + uncentered RePCB similarity (ours)1208.547.12.781.7360.873

The topology-sensitivity study below uses 30 layout pairs per category, rated by three senior PCB engineers and GPT-6 Astra. RePCB scores are higher for topologically similar pairs and lower for topologically different pairs despite their similar WL/via values. This is evidence of topology sensitivity on these selected pairs, rather than an independent preference or quality assessment:

Type∣ΔWL∣\vert\Delta\mathrm{WL}\vert(%)∣ΔVia∣\vert\Delta\mathrm{Via}\vert(%)Expert Topology Similarity ↑\uparrowGPT-6 Astra Topology Similarity ↑\uparrowRePCB Similarity ↑\uparrow
Different WL/Via, Similar Topology38.5±13.126.4±8.80.91±0.040.93±0.020.904±0.031
Similar WL/Via, Different Topology1.7±0.91.9±0.70.26±0.060.23±0.040.312±0.067

2. Surrogate Rewards and Physical Performance

RePCB is supervised by WL and via count, not EM or thermal labels. Its role during specialization is to guide layouts toward expert-reference outcomes; direct physical refinement is performed by MM-Evolve. The reward ablation above shows that replacing the learned reward with analytic metric similarity changes the final EM/thermal costs from 2.78/1.736 to 4.87/2.214. This supports the learned reward's utility in the full pipeline, without treating similarity as a direct physical objective. Instead, it captures PnR design topologies from the PnR topologies of expert-designed PCBs. The component ablation below evaluates SFT, DAPO, RePCB, and MM-Evolve:

ImplementationWL(mm) ↓\downarrowVia ↓\downarrowSEM↓\mathcal{S}_{\mathrm{EM}}\downarrowSTh↓\mathcal{S}_{\mathrm{Th}}\downarrowRePCB similarity to Expert Designs ↑\uparrow
w/o SFT1453.565.65.762.1870.597
w/o DAPO1389.661.46.212.2470.642
w/o RePCB1176.950.37.062.5840.541
w/o MM-Evolve1219.252.28.742.5780.824
EvoPCB1208.547.12.781.7360.873

The w/o RePCB row corresponds to the “WL+Via only” configuration in the reward ablation above. The gap between “w/o MM-Evolve” and EvoPCB supports the contribution of the MM-Evolve stage, which combines visual refinement and direct simulation feedback. The SFT/DAPO/RePCB initialization variants use the same MM-Evolve setting; their remaining gaps suggest a contribution from initialization.

3. Computational Efficiency

The runtime measurements below report per-board design time averaged over the 11 benchmark cases. EvoPCB takes 1045.89s, comprising approximately 301.8s for the initial stage and 744.1s for MM-Evolve, including physical refinement. MM-Evolve averages 3.7 EM and 2.4 thermal calls per board. The operation-level measurements are 131.5s per EM simulation, 13.8s per thermal simulation, 31.2s per multimodal analysis, and 20.5s per script-generation call. Thus, refinement includes analysis and generation as well as physics simulation.

In the reported comparison, EvoPCB is faster than NS-Place, PCBAgent, DeepPCB, and GPT-6 Astra. PCB-PR-App is faster but has worse averages on the four quality metrics. These runtime comparisons include EvoPCB's physical refinement but do not by themselves establish matched physical-evaluation budgets across methods.

CaseEM callsThermal callsMM-Evolve time (s)
P153996.4
P232599.7
P332601.2
P453995.1
P543805.6
P622408.9
P742790.3
P832598.8
P932600.5
P1053997.0
P1142791.6
Average3.72.4744.1
MethodTime(s)↓\downarrowRatio
PCB-PR-App344.90.33×
NS-Place+FreeRouting8682.78.30×
PCBAgent+FreeRouting1788.21.71×
DeepPCB7378.37.05×
GPT-6 Astra2620.02.51×
EvoPCB1045.891.00×

4. Dataset Distribution and Generalization Analysis

The dataset distribution below reports component, net, and layer counts for the 2341 SFT and 5741 DAPO layouts sourced from an industrial company. The paired scripts are generated by constrained BO. None of the 11 benchmark boards or the below three larger cases appears in the training data.

Subset≤40 comps41–80 comps>80 comps≤50 nets51–100 nets>100 nets2 layers4–6 layers>6 layers
SFT Dataset615129842813059201161492665184
DAPO Dataset1392317611733025238233435361650555

The larger-board evaluation below uses three held-out boards with 124–236 components and 4–8 layers, larger than the original test cases. EvoPCB has lower WL, via count, EM cost, thermal cost, and runtime than DeepPCB, the powerful commercial tool, in all three cases. This supports transfer to these larger unseen boards.

Case#Comp#Pins#Nets#LayersMethodWL(mm) ↓\downarrowVia ↓\downarrowSEM↓\mathcal{S}_{\mathrm{EM}}\downarrowSTh↓\mathcal{S}_{\mathrm{Th}}\downarrowRuntime(s) ↓\downarrow
Large-1124386964DeepPCB4976.21627.312.66410532.6
EvoPCB4528.71463.842.1261028.4
Large-21685241316DeepPCB7598.62478.423.08612374.1
EvoPCB6781.32184.272.3811512.7
Large-32368112058DeepPCB16192.54919.173.51119016.3
EvoPCB10644.83475.062.7142386.9

Seed 2026 is the default setting used for the original table. To assess run-to-run variability separately, we additionally evaluated 30 random seeds and report the mean, standard deviation, and range as follows:

StatisticWL(mm) ↓\downarrowVia ↓\downarrowSEM↓\mathcal{S}_{\mathrm{EM}}\downarrowSTh↓\mathcal{S}_{\mathrm{Th}}\downarrowTime(s) ↓\downarrow
EvoPCB (default)1208.547.12.781.7361045.89
mean±std1214.8±18.747.6±1.82.91±0.421.748±0.0281048.7±22.9
min/max1187.9/1248.645.1/50.82.24/3.631.704/1.7971012.4/1088.3

In the 30-seed evaluation reported here, we keep the trained checkpoint, benchmark boards, objective weights, iteration limits, and simulator settings fixed, and vary only the seeds of the stochastic EDA operations, namely simulated-annealing placement and rip-up-and-reroute. Different seeds can therefore produce different placement trajectories and routing solutions, which may lead MM-Evolve along different refinement paths and cause variation in WL, via count, EM/thermal costs, and runtime. The EM and thermal solvers themselves are deterministic for a fixed layout and configuration, so their variation is induced indirectly by the resulting layouts.

The 30-seed results further show that even the largest observed EvoPCB averages for WL (1248.6), EM (3.63), and thermal cost (1.797) are below the corresponding best reported baseline averages (1346.1, 7.30, and 2.001). Via count is less conclusive: its range is 45.1–50.8, compared with DeepPCB's reported 48.8. These observations support robustness to the evaluated EvoPCB seeds, without accounting for baseline run-to-run variability.

Conclusion

Thank you again for your valuable and constructive feedback. Your comments have substantially improved the clarity of our paper. We would be grateful if you could reevaluate our work and reconsider the score in light of these clarifications and additional results.

Dear Reviewer,

We hope our clarification has addressed your concerns. We have added further experiments to the anonymous repository: https://anonymous.4open.science/r/EvoPCB-ICLR-2027-8E0E/Rebuttal_EvoPCB.md. We welcome any feedback you may have, and we would be very grateful if you could reconsider your score in light of our clarification.

Best regards,

All authors

Comments on the paper

1 comment

Dear Reviewer,

As the discussion phase draws to a close, please acknowledge the author's response and let them know whether it has addressed some or all of your concerns.

Thank you for the time and effort you and the other reviewers have put into this discussion.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.