acceptodds
Under review as a conference paper at ICLR 2027

CADET: Construction of PCB Routing Instances with Executable Demonstrations

Abstract

Printed circuit board (PCB) routing is the stage of physical layout that finds paths connecting the placed pads of each net. Public datasets in this domain are few, and most contain only the final states of routed boards, not demonstrations of the electronic design automation (EDA) engine actions that produce them. To close this gap, we propose CADET (Constructing executAble Demonstrations for Engine-grounded Training), which jointly generates synthetic boards and demonstrations built from EDA engine actions. The resulting RouteDemos provides 100,000 synthetic boards routed with zero design rule violations (DRVs), together with their demonstrations. Their distributions over 17 statistics resemble those of 142 real reference boards. CADET trains a routing policy on these demonstrations with behavior cloning and fine-tunes it with reinforcement learning (RL). On real test boards, the policy raised the rate of complete DRV-free routing from 38.7% to 66.7% over an RL-only baseline and exceeded all other baselines in this rate, including open-source autorouters and GPT-6 Astra. We release the dataset and its generation code as open source.

Primary area:
datasets and benchmarks
Keywords:
PCB routing, electronic design automation, synthetic data generation, routing demonstrations, behavior cloning, reinforcement learning

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

CADET builds executable PCB routing demonstrations using native EDA operations, then uses them for BC and PPO training.

Critical strengths

I think the dataset is useful, and the replay checks are a good part of the work. The BC-only and full-pipeline results also show that these demonstrations can help with real-board routing.

Critical weaknesses
  1. I'm not quite convinced by the explanation in Section 4.4. The authors describe demonstrations as the main source of the CP gain, then qualify this by saying the comparison covers the whole training procedure. That qualification still doesn't tell me where most of the gain comes from. Section 3.3 also changes the KL penalty and critic training, so I would be more cautious with the claim about demonstrations having a dominant effect.

  2. The gains over the learning baselines look good, but I'm less sure how stable they are after retraining. The paper reports only one PPO training seed. Repeating evaluation on the trained policy, as described in Section 4.1, doesn't cover variation from training itself. I would like to see the main comparison repeated with different seeds under same settings.

Questions for authors

What evidence makes the authors think demonstrations are the main source of the CP improvement, given that the KL and critic settings also change?

2 comments
AuthorsOpenReview ↗

Response to Reviewer bzJB

Thank you for the two constructive requests.

 

W1 / Q1. Source of the CP gain

W1: "...Section 3.3 also changes the KL penalty and critic training, so I would be more cautious with the claim about demonstrations having a dominant effect."

Q1: "What evidence makes the authors think demonstrations are the main source of the CP improvement, given that the KL and critic settings also change?"

We agree that §4.4 did not separate these. Starting from each seed's BC checkpoint, we reran RL fine-tuning with one setting or both turned off (D3-B test. CP: mean ± std over the 4 seeds).

MethodBC trainingRL trainingβCritic detachedCPΔΦRout.
PPO (RouteDemos)✗✓–✗38.0 ± 6.296.092.0
CADET (BC)✓✗––43.1 ± 2.178.381.7
CADET✓✓0.05✓67.2 ± 1.8107.291.7
CADET✓✓0✓64.0 ± 3.4109.190.1
CADET✓✓0.05✗66.5 ± 1.5107.791.0
CADET✓✓0✗63.3 ± 2.9109.188.7

All four CADET settings reached 63.3–67.2% CP, more than 25 pp above PPO (RouteDemos) (38.0%). So the two settings (KL penalty, critic detachment) are not the main source of the gain. The main source is BC on demonstrations followed by RL fine-tuning (43.1% with BC alone, 67.2% after fine-tuning) (Revised: §4.4, App. G).

Critic warm-up is needed because BC trains only the policy. Without it, the first PPO advantages would come from an untrained critic. The actor is frozen during warm-up, so RL starts from the BC policy unchanged (App. B), and warm-up counts toward the 200 fine-tuning iterations (Revised: App. I).

 

W2. Training seeds

"The gains over the learning baselines look good, but I'm less sure how stable they are after retraining. The paper reports only one PPO training seed. ... I would like to see the main comparison repeated with different seeds under same settings."

We retrained the full pipeline with 4 training seeds (42, 43, 44, 45). CADET was trained with BC and RL fine-tuning, and each learned baseline with scratch PPO, and each seed's selected checkpoint was evaluated with the paper's protocol (4 runs) (Revised: Table 2, 3; std in Table 9).

MethodD3-B test (Submitted)D3-B test (Revised)D3-A (Submitted)D3-A (Revised)
CADET @166.767.2 ± 1.891.489.5 ± 1.0
CADET @380.379.8 ± 0.397.596.3 ± 0.3
CADET (BC)49.343.1 ± 2.183.383.2 ± 0.8
PPO (RouteDemos)38.738.0 ± 6.279.580.2 ± 4.7
PPO (D2)37.440.4 ± 2.275.378.4 ± 3.3
PPO (D3-B ref)33.238.3 ± 2.074.075.6 ± 1.6

CP (%). Submitted: a single training run. Revised: mean ± std over the 4 seeds.

On both D3-B test and D3-A, CADET@1's CP was higher than every PPO row, and the 95% CIs did not overlap (Revised: App. G; New: Table 11).

Your two requests let us analyse the source of the CP gain more rigorously. We would be glad to answer any further questions.

Hi, thank you for the response and the additional experiments. The four-seed results solved my concern about training stability. The fine-tuning ablations also show that the KL penalty and critic detachment do not explain the main gap over PPO. Section 4.4 reads much more clearly now. I can see what BC alone achieves and how much RL fine-tuning adds. As my 2 concerns have been resolved, I will raise my rating from 2 to 3.

Reviewer_ayrrOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper propose a RL-based method called CADET to execuate routing in PCBs along with a dataset RouteDemos. Experimental results show effectiveness of proposed method.

Critical strengths

1.The paper and method proposed are of high level completion; 2.The proposed dataset seem to align with the real distribution of PCBs in size, density and design rules; 3.The experiments show some level of effectiveness.

Critical weaknesses

1.The training is limited to to double-layer boards, testing does not exceed one hundred solder pads, and lacks large-scale generalization validation; 2.The single success rate of D3-B is only slightly higher than that of traditional routers, and the connection rate and number of vias are still poor; 3.The seed used in training is single.

Questions for authors

see in weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer ayrr

Thank you for the clear comments.

W1. Evaluation scope: two-layer boards, up to 100 pads

"The training is limited to to double-layer boards, testing does not exceed one hundred solder pads, and lacks large-scale generalization validation"

We agree that boards with hundreds of pads or more than two layers remain untested, and we added a Limitations section that states this (New: §6 Limitations). D3-B (31–100 pads) has 283 two-layer boards out of 286, and we used all of them: 142 to fit the generator and 141 for testing (§4.1). The remaining three multilayer boards were too few for the generator to reproduce a multilayer distribution.

 

W2. CADET@1 vs. FreeRouting

"The single success rate of D3-B is only slightly higher than that of traditional routers, and the connection rate and number of vias are still poor"

We agree. Compared with FreeRouting (45.6 s), CADET@1 (15.5 s per board) had comparable CP (67.2% vs. 65.6%) and worse routability and via count (91.7% vs. 96.8%; 10.4 vs. 4.1). With three rollouts in a similar total time (47.6 s), CADET@3 had higher CP (79.8%) and comparable routability (95.9%), and still used more vias (9.1) (Revised: §4.3, Table 9; New: §6 Limitations).

The larger number of vias appears to come from the BC stage. The BC-only policy used far more vias than PPO (RouteDemos), which was trained by RL alone. RL fine-tuning lowered the count, and removing the KL penalty that keeps the policy close to the BC policy (β = 0) lowered it further (Revised: §4.4, Table 9; New: Table 10).

MethodCPVias
FreeRouting65.64.1
PPO (RouteDemos)38.04.8
CADET (BC)43.111.9
CADET@1 (β = 0.05)67.210.4
CADET@1 (β = 0)64.07.7

 

W3. Training seeds

"The seed used in training is single."

We retrained the full pipeline with 4 training seeds (42, 43, 44, 45) and evaluated each seed's selected checkpoint with the paper's protocol (4 runs) (Revised: Table 2, 3; std in Table 9).

MethodD3-B test (Submitted)D3-B test (Revised)D3-A (Submitted)D3-A (Revised)
CADET @166.767.2 ± 1.891.489.5 ± 1.0
CADET @380.379.8 ± 0.397.596.3 ± 0.3
CADET (BC)49.343.1 ± 2.183.383.2 ± 0.8
PPO (RouteDemos)38.738.0 ± 6.279.580.2 ± 4.7
PPO (D2)37.440.4 ± 2.275.378.4 ± 3.3
PPO (D3-B ref)33.238.3 ± 2.074.075.6 ± 1.6

CP (%). Submitted: a single training run. Revised: mean ± std over the 4 seeds.

We would be glad to discuss anything further.

Thanks for your additional experiments, the rating will be kept.

Reviewer_5wPUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces CADET (Constructing executAble Demonstrations for Engine-grounded Training), a framework for generating synthetic PCB routing tasks with executable demonstrations. Instead of relying on existing routing datasets, the authors generate routable boards directly in KiCad by incrementally adding pads and retaining only those for which a valid, design-rule-clean routing solution can be found. They use best-first search to generate routing actions & replay each demonstration in a fresh engine instance to verify its correctness. RouteDemos dataset contains 100k training boards & 128 validation boards, with its characteristics calibrated against 142 real boards. The authors train a Transformer-based routing policy using behavior cloning followed by PPO. Using evaluation on real test boards, CADET achieves a 66.7% clean-pass rate, outperforming the RL-only baseline, as well as the evaluated rule-based autorouters and LLM-based agents.

Critical strengths
  • The paper is clear and easy to understand.
  • The authors address an important limitation w.r.t the data in PCB routing by generating synthetic boards along with executable demonstrations of the routing process using native EDA actions.
  • The fresh-engine replay check is a good design choice, as it ensures that the demonstrations avoid design-rule violations.
  • The dataset is valuable & the authors show that its board statistics are reasonably close to those of 142 real reference boards.
  • The results on real test boards are promising, with clear improvements over the reported baselines.
Critical weaknesses
  • Though the synthetic data is based on 142 real reference boards, it may not necessarily capture the structural relationships and complexity of real PCB designs.
  • The evaluation is limited to 2-layer boards, so it still remains unclear whether the approach can handle multilayer and more complex industrial PCB designs.
  • Though the demonstrations guarantee valid, route avoiding design rule violations, the authors don't show whether these routes are of high quality/near-optimal routing solutions.
  • The authors report averages and standard deviations across 4 runs, but don't provide confidence intervals or statistical significance tests to support the reported improvements.
Questions for authors
  1. It would be to account for the computational cost of generating and validating RouteDemos, including how this cost compares with the training cost of the RL-only baseline. How do they compare?
  2. How do the generated routing demonstrations compare with human designed routes in terms of wirelength/overall routing quality?
  3. Can the authors report results across multiple independent training runs and show the variance in clean-pass rates between BC+RL and scratch PPO?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 5wPU (1/2)

Thank you for the thoughtful review. We answer each point below.

W1. Do synthetic boards capture real structure?

"Though the synthetic data is based on 142 real reference boards, it may not necessarily capture the structural relationships and complexity of real PCB designs."

We agree that RouteDemos may not reproduce all of the complex structure of real PCBs. Measuring distributional similarity between high-dimensional datasets is a long-standing challenge. The paper assessed it from two complementary angles: distances between marginal distributions and performance on a downstream task.

First, distances between statistic distributions (§4.2, Table 1). For 17 main board statistics (e.g. board size, pad density), we computed the Kolmogorov–Smirnov (KS) distance between the marginal distribution of each dataset and that of the target distribution, D3-B ref. The mean KS distances were 0.064 / 0.086 / 0.773 for D3-B test / RouteDemos / D2, the existing synthetic dataset of PCBWorld (Song et al., 2026). In KS distance, RouteDemos was much closer to the target distribution (D3-B ref) than D2, and at a level similar to D3-B test, another split of the same distribution.

Second, how well a policy trained on each dataset solves unseen real boards (Revised: §4.4, Table 3). We trained policies with the same method on different training boards and evaluated them on the real D3-B test boards.

MethodTraining boardsCPΔΦRout.
PPO (D2)existing synthetic (PCBWorld)40.480.6674.4
PPO (D3-B ref)real boards (D3-B ref)38.394.8791.4
PPO (RouteDemos)RouteDemos38.095.9892.0

RouteDemos was built from the attribute distributions of D3-B ref only (App. K), and D3-B test was used neither in generation nor in training (§4.1, App. B, App. H). At a similar or slightly lower CP (38% against 40%), PPO (RouteDemos) and PPO (D3-B ref) had much higher ΔΦ and routability than PPO (D2). PPO (RouteDemos) was on par with PPO (D3-B ref), which was trained on real boards. We revised the training-board comparison of §4.4 to this conclusion (Revised: §4.4, Table 3).

 

W2. Evaluation scope: two-layer boards

"The evaluation is limited to 2-layer boards, so it still remains unclear whether the approach can handle multilayer and more complex industrial PCB designs."

We agree that multilayer and more complex industrial boards remain untested, and we added a Limitations section that states this (New: §6 Limitations). D3-B (31–100 pads), the public benchmark we target, has 283 two-layer boards out of 286, and we used all 283 (§4.1). The remaining three multilayer boards were too few for the generator to reproduce a multilayer distribution.

 

W3 / Q2. Demonstration quality against human routes

W3: "...the authors don't show whether these routes are of high quality/near-optimal routing solutions."

Q2: "How do the generated routing demonstrations compare with human designed routes in terms of wirelength/overall routing quality?"

Each demonstration is not an optimal routing but one that is guaranteed to be a clean pass (App. K). To evaluate the routing quality of the demonstrations, we re-routed from scratch the 142 D3-B ref boards, which have human routes. We used the route-generation procedure of the demonstrations (TryConnect of Algorithm 2 and the repair of Algorithm 3). Pad positions are fixed here, so unlike in RouteDemos, DRVs or routing failures can occur. The search completed 88 boards with a clean pass (same DRC as Table 9), and we compared wirelength and vias on these boards.

MethodWL (mm)Vias
Human reference5403.9
Demonstration search5536.0

On these boards, the demonstration wirelength was 1.02× the human layout's (per-board median), and the demonstrations used 2.1 more vias on average. These boards are likely the easier ones (New: App. L).

AuthorsOpenReview ↗

Response to Reviewer 5wPU (2/2)

W4 / Q3. Training seeds and statistics

W4: "The authors report averages and standard deviations across 4 runs, but don't provide confidence intervals or statistical significance tests to support the reported improvements." Q3: "Can the authors report results across multiple independent training runs and show the variance in clean-pass rates between BC+RL and scratch PPO?"

We retrained the full pipeline with 4 training seeds (42, 43, 44, 45). CADET was trained with BC and fine-tuning, and each baseline with scratch PPO, with 4 evaluation runs per seed (Revised: Table 2, 3, 9, Fig. 8; New: Table 11).

MethodD3-B test (Submitted)D3-B test (Revised)D3-B test (95% CI)D3-A (Submitted)D3-A (Revised)D3-A (95% CI)
CADET @166.767.2[64.3, 70.1]91.489.5[87.9, 91.1]
CADET @380.379.8[79.4, 80.3]97.596.3[95.8, 96.8]
CADET (BC)49.343.1[39.8, 46.4]83.383.2[82.0, 84.5]
PPO (RouteDemos)38.738.0[28.2, 47.8]79.580.2[72.7, 87.7]
PPO (D2)37.440.4[36.9, 43.9]75.378.4[73.2, 83.6]
PPO (D3-B ref)33.238.3[35.1, 41.4]74.075.6[73.1, 78.1]
PPO (RouteDemos w/o TryConnect)34.030.9[23.5, 38.2]76.376.6[74.9, 78.4]

CP (%). Submitted: a single training run. Revised: mean over the 4 seeds. 95% CI: t interval of that mean.

On both D3-B test and D3-A, CADET@1's CP was higher than every PPO row, and the 95% CIs did not overlap (Revised: App. G).

 

Q1. Compute cost

"...account for the computational cost of generating and validating RouteDemos, including how this cost compares with the training cost of the RL-only baseline."

MethodTraining dataRL iterData generation (CPU only)BC training (RTX PRO 6000 ×4)RL training (L40 ×1)
PPOD3-B ref500––177 h
PPORouteDemos500146 h–135 h
CADETRouteDemos200146 h97 h85 h

Times are approximate. The data-generation time includes the replay check and is for a single server (App. A.2). Training times are for a single run. The RL-only baseline is the two PPO rows (Revised: App. I; New: Table 13).

We hope the additional experiments (the 4-seed runs and the comparison with human layouts) and the revisions answer your questions. We would be glad to answer anything further during the discussion period.

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,

Your AC

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank the three reviewers for their careful reading, insightful comments and constructive feedback. All reviewers recognised the value of the RouteDemos dataset and that training with it improved routing performance on real PCBs [5wPU, ayrr, bzJB]. We are also grateful that the two core goals of this work were recognised: executable demonstrations verified by fresh-engine replay [5wPU, bzJB] and a board distribution close to that of real PCBs [5wPU, ayrr]. We also appreciate the positive view of the paper's clarity [5wPU, ayrr] and completeness [ayrr]. Our individual responses are in the Official Comments under each review. In the responses, (Revised: …) marks a revised place and (New: …) a new one; a location without a label, e.g. (§4.1), was already in the paper.

The revised PDF includes the following changes.

  1. Training seeds [5wPU, ayrr, bzJB]. We retrained the full pipeline with 4 training seeds (42, 43, 44, 45). The results are now means over the 4 seeds (Revised: Table 2, 3, Fig. 8), with the std (Revised: Table 9) and 95% CIs for CP (New: Table 11), and the text follows the new results (Revised: §4.1, §4.4, App. B, App. G, App. H).
  2. Limitations [5wPU, ayrr]. We added a Limitations section. It states that boards with hundreds of pads or more than two layers remain untested, and that CADET@1 had lower routability and more vias than FreeRouting (New: §6 Limitations).
  3. Source of the CP gain [bzJB]. We added an ablation that turns off one fine-tuning setting or both (KL penalty, critic detachment) (Revised: §4.4, App. G).
  4. Comparison with FreeRouting [ayrr]. We toned down the claim of CADET's superiority in CP (Revised: abstract, §1, §4.3, §7). We stated that CP was comparable to FreeRouting in less time (one rollout) and higher at a similar time (three rollouts) (Revised: §1, §4.3, §7).
  5. Demonstration quality [5wPU]. We compared the routing quality of the demonstration search with human layouts (New: App. L).
  6. Compute cost [5wPU]. We added a comparison of the generation, BC and fine-tuning costs with the training cost of scratch PPO (Revised: App. I; New: Table 13).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.