acceptodds
Under review as a conference paper at ICLR 2027

PlaceReasoner-Pro: An 8B Open-Weight Reasoning Model for Frontier-Level VLSI Macro Placement

Abstract

Reasoning-driven agents have substantially advanced VLSI (i.e., chip) macro placement, but their reliance on proprietary frontier models raises severe concerns on confidentiality, specialization, and inference cost. We present PlaceReasoner-Pro, an 8B open-weight vision-language model for macro placement, trained with supervision that moves step by step from cheap geometric checks to the true objective: post-route power, performance, and area (PPA). Because real designs are proprietary and each place-and-route (P&R) run is expensive, we first build a circuit-agnostic synthetic environment that generates realistic floorplans (core outlines, macro sizes, and pin sides) and automatically verifies that a placement is complete, inside the core, and overlap-free; it yields 10K verified, reasoning-annotated placements for supervised fine-tuning. Two-stage reinforcement learning then refines the model: 1) Stage-I uses this verifier as the reward to make placements geometrically legal, raising overlap-free placement from 78.1% to 90.9%; and 2) Stage-II runs every candidate through a full OpenROAD P&R flow and uses the measured PPA as the reward, improving total negative slack (TNS) over Stage-I on 13 of 16 benchmark tasks. On these 16 tasks, PlaceReasoner-Pro achieves timing comparable to Claude Opus 4.8 in the same harness (7 vs. 7 wins on worst negative slack, 7 vs. 9 on TNS) while running entirely on premises. Our results show that compact, domain-specialized small models can approach frontier-level macro placement without relying on proprietary inference.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
VLSI, macro placement, agentic, RL, dataset, PPA, Layout

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_pQ6hOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents an 8B open-weight vision-language model (VLM) for macro placement in VLSI named PlaceReasoner-Pro, to introduce a performance-wise comparable alternative to the proprietary frontier models. This model is trained in multiple supervision stages, ranging from geometric checks to PPA optimization. The authors describe their supervision stages in detail, thereby introducing a synthetic placement environment and associated 10K-SFT/2K-RL training corpus that teaches transferable geometric placement reasoning without proprietary design data. This enables them to minimize costly use of OpenROAD for EDA during the reinforcement learning stages. The paper's experiment section evaluates the training results stage-wise, isolating the contribution of each stage. Additionally, they compare their model against Claude Opus 4.8 using the same harness and achieve comparable PPA results.

Critical strengths
  • They target the major problem of confidentiality issues raised by the use of proprietary frontier models, which are externally hosted, by presenting a comparable alternative based on an open-weight model.

  • By creating a synthetic environment for their first reinforcement learning stage, they can largely reduce the computational effort compared to a full EDA environment. This allows them to train the model on the geometric checks much faster.

Critical weaknesses
  • After the paper highlights the monetary cost of using proprietary frontier models as a weakness, it does not evaluate its model costs, such as its cost per design.
Questions for authors
  • Can you provide any details about the average power consumption or cost per design?
2 comments
AuthorsOpenReview ↗

Cost per design: measured GPU energy and per-task cost (App. A.8)

Q1 / W1: Cost per design.

Thank you for the positive assessment and for raising the cost question. We agree that cost should be reported alongside our critique of proprietary models, and we have added this analysis to the revised paper ([Appendix A.8]).

Summary. PlaceReasoner-Pro uses 0.0844 kWh of GPU energy per task on average. At the U.S. average commercial electricity rate (~0.145 USD/kWh), this costs about 0.01 USD in GPU electricity (about 0.025 USD including the idle draw of a dedicated GPU). By comparison, frontier-model placement has been reported to cost about 310 USD per design in API tokens (Kahng et al., 2026). PlaceReasoner-Pro therefore runs entirely on-premises at a small fraction of that cost, while achieving timing comparable to Claude Opus 4.8 under the same harness.

Measurement. Hardware: 8× NVIDIA RTX 6000 Ada (48 GB, 300 W cap), the same server used for all PlaceReasoner-Bench runs. The model is served by vLLM 0.26 with one full BF16 replica per GPU, so each request runs on one GPU. We calibrated on 18 single-stream requests (3 repeats of 6 types: 512/1024-token decode, 8K/16K text prefill, image+text prefill, image + 8K text + 512-token decode). Every prompt and image was unique, so there were no cache hits. Each request was attributed to the GPU whose vLLM token counters rose by exactly that request's usage, and that GPU's instantaneous power (10 Hz) was integrated above its idle baseline over the request window plus a 2.5 s tail. We fit E = e_p · prompt + e_c · completion (least squares, no intercept).

QuantityValue
e_p, energy per prompt token (text or image)0.0585 ± 0.0022 J
e_c, energy per completion token5.747 ± 0.038 J
Fit R² (18 requests)0.9985
Power while decoding (1 GPU)286 W
Decode speed (single stream)48.4 tok/s
Idle power per GPU (model loaded)25.2 W

± is the standard error of the fit. A completion token costs about 98× a prompt token: prefill is compute-bound and batched over the prompt, while decoding is memory-bandwidth-bound (each token reads all 17.5 GB of weights, so the GPU runs near its 300 W cap at ~48 tok/s).

Energy per task. LLM energy = e_p · prompt + e_c · completion, the GPU energy that model calls add above idle. + Idle GPU adds one GPU's idle draw over the task's full wall-clock runtime, i.e., the cost of reserving a dedicated GPU, including the time it waits while OpenROAD runs.

TaskPrompt tokCompletion tokLLM energy (kWh)Runtime (h)+ Idle GPU (kWh)
ariane133 1:1336,82270,8690.11862.460.1807
ariane133 2:1475,13276,3200.12953.410.2154
ariane81 1:1486,69078,8270.13372.900.2070
ariane81 2:1661,533101,3770.172611.390.4598
bp_be 1:178,0378,5830.01500.780.0347
bp_be 2:1313,32231,2630.055012.980.3821
bp_fe 1:1308,40231,8500.05590.700.0734
bp_fe 2:1154,08610,5480.01930.560.0335
ethernet 1:1340,64848,6880.08330.360.0923
ethernet 2:1587,44168,3290.11860.480.1308
swerv_wrapper 1:1250,30237,6320.06411.460.1009
swerv_wrapper 2:1397,87046,3610.08059.030.3082
vga_lcd 1:1388,76163,6410.10791.590.1480
vga_lcd 2:1483,36159,1290.10225.360.2373
VeriGPU 1:1329,73525,9430.04681.450.0834
VeriGPU 2:1381,02625,9110.04750.950.0716
Total (16 tasks)5,973,168785,2711.350555.902.7591
Mean per task373,32349,0790.08443.490.1724

The whole 16-task suite uses 1.351 kWh of LLM GPU energy (fit standard error ±0.008 kWh). Completion tokens are 12% of the tokens but 93% of the energy.

We hope these responses address your concerns, and we are happy to answer any further questions during the discussion period.

Reviewer_pQ6hOpenReview ↗

Thanks

Thanks for replying to my queries and explaining further. My rating can hence be maintained as is.

Reviewer_69ehOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PlaceReasoner-Pro, an 8B open-weight VLM specialized for VLSI macro placement. The training pipeline progressively increases supervision fidelity: synthetic SFT on verified placement trajectories, Stage-I RL with inexpensive geometric rewards, and Stage-II RL using full OpenROAD P&R outcomes including WNS, TNS, DRC, power, and area. The paper argues that domain-specific post-training can enable a compact open model to approach the macro-placement quality of Claude Opus 4.8 while allowing fully on-premises deployment.

Critical strengths
  1. The progressive use of increasingly faithful supervision is practically meaningful. The paper separates inexpensive geometric competence learning from costly physical-design optimization, which is well motivated for macro placement. In particular, Stage-II evaluates candidates through standard-cell placement, CTS, global and detailed routing, and timing repair before computing reward, making the training objective considerably closer to actual downstream design quality than HPWL-style proxies.
  2. The empirical gains of the specialized 8B model are substantial. Stage-I significantly improves geometric validity, while Stage-II improves WNS and TNS on 13 of 16 evaluation tasks relative to Stage-I. Under the same PlaceReasoner-Beta harness, the final model also achieves timing results close to Claude Opus 4.8, suggesting that domain-specific post-training can substantially reduce the gap between compact open-weight and frontier proprietary models.
Critical weaknesses
  1. The “frontier-level” comparison to Claude Opus 4.8 lacks statistical robustness. The main comparison is based on 8 designs under two aspect ratios, i.e., 16 tasks, and is summarized largely through win counts such as 7 vs. 9 on TNS and 7 vs. 7 on WNS. The paper does not report repeated independent generations, variance, or confidence intervals for these stochastic VLM policies. As a result, it is difficult to determine whether the observed parity reflects a robust model-level result or particular sampled placements. This directly affects one of the paper’s headline claims.
  2. The generalization and progressive-training claims are not cleanly isolated. Although the exact Stage-II training tasks are disjoint from the evaluation tasks, the training and evaluation sets contain related design families such as Ariane and BlackParrot, and both are dominated by processor-style designs. The authors themselves note that the train and test sets are close in design class. Therefore, the current evidence is stronger for within-domain transfer than for broad cross-design generalization. Moreover, the paper does not compare SFT → Stage-II RL against SFT → Stage-I RL → Stage-II RL under the same P&R evaluation budget, so the claimed benefit of the progressive-fidelity curriculum itself is not directly established.
  3. The methodological novelty is limited. The individual learning components are largely standard: supervised fine-tuning with LoRA, GRPO/DAPO-style RL, verifier-based rewards, and a second RL stage using a more expensive environment. The paper’s own contribution summary primarily emphasizes the synthetic environment, dataset, progressive training setup, and controlled evaluation. These are valuable system and domain contributions, but the paper does not introduce a substantially new learning objective, architecture, or reasoning mechanism. Stronger evidence is needed that the proposed synthetic-to-physical curriculum represents a generalizable learning principle rather than an effective EDA-specific application of a standard small-model post-training recipe.
Questions for authors
  1. Can the authors evaluate PlaceReasoner-Pro-II and Claude Opus 4.8 with multiple independent generations per task and report mean/median performance, variance, and confidence intervals? Does the “comparable frontier-level performance” conclusion remain unchanged?
  2. Under the same number of P&R evaluations, how does direct SFT → Stage-II RL compare with SFT → Stage-I RL → Stage-II RL? This experiment would directly test whether Stage-I improves the sample efficiency of expensive physical-feedback training.
  3. Can the authors provide a stricter leave-design-family-out evaluation, e.g., excluding Ariane/BlackParrot-related designs from Stage-II training and evaluating on those families afterward? This would more clearly establish cross-design transfer.
  4. Stage-II improves timing but reduces the number of DRC-clean layouts from 11/16 to 8/16. Would it be more accurate to characterize the result as a timing improvement with a routability trade-off rather than a general PPA improvement?
4 comments
AuthorsOpenReview ↗

Response (1/2): Multi-run Opus comparison, leave-family-out generalization, and scoped contributions

We thank the reviewer for the careful reading and constructive questions. We address each point below; all new results have been added to the revised paper, highlighted in blue.

W1/Q1: Statistical robustness

We ran 3 independent generations per task for both PR-Pro-II and PR-Beta (Opus 4.8) under the identical harness and P&R flow (App. A.6, Table 7). Mean WNS/TNS (ns):

TaskWNS PR-Pro-IIWNS OpusTNS PR-Pro-IITNS Opus
ariane133 1:1-0.030-0.190-1.25-234.0
ariane133 2:1-0.020-0.058-4.05-11.6
ariane81 1:1-0.030-0.066-0.21-12.7
ariane81 2:1-0.127-0.172-0.92-172.2
bp_be 1:1-0.210-0.226-14.67-19.3
bp_be 2:1-0.237-0.277-21.03-25.7
bp_fe 1:1-0.073-0.257-0.95-15.9
bp_fe 2:1-0.097-0.152-0.28-2.5
ethernet 1:1-0.277-0.314-1.34-1.5
ethernet 2:1-0.273-0.310-1.33-1.5
swerv_wrapper 1:1-0.533-0.522-504.5-480.6
swerv_wrapper 2:1-0.570-0.586-477.8-534.0
vga_lcd 1:1-0.717-0.800-39.6-96.0
vga_lcd 2:1-0.683-0.830-27.5-188.5
VeriGPU 1:1-0.260-0.260-13.19-12.6
VeriGPU 2:1-0.273-0.274-13.73-26.6

PR-Pro-II has the better mean WNS on 14/16 tasks (1 tie) and the better mean TNS on 14/16. We attribute this mainly to consistency. Stage-II RL trains the policy on routed PPA from real designs, so it learns which placement styles and halo/corridor/spacing settings pay off after routing and applies them reliably. Opus 4.8 has no such feedback and occasionally produces poor placements that dominate its mean (e.g., TNS −234.0 ± 330.4 ns on ariane133 1:1 and −172.2 ± 327.9 ns on ariane81 2:1). With n=3, the 95% CIs are wide and overlap on most tasks, so we read this conservatively: these results do not establish that PR-Pro-II is superior, but they confirm that the comparable placement quality we report is not an artifact of a single sample.

W2/Q2: Under the same number of P&R evaluations, how does direct SFT → Stage-II RL compare with SFT → Stage-I RL → Stage-II RL?

Under the same number of P&R evaluations, starting Stage-II directly from SFT would waste much of the expensive P&R budget, because the SFT model is still weak in placement quality and principle alignment; Stage-I fixes this at negligible cost.

Stage-I uses no P&R evaluations (its verifier runs in milliseconds), and relative to SFT it raises overlap-free placement from 78.1% to 90.9% and containment from 84.3% to 92.9%, cuts row overflow from 0.86 to 0.22 per task, and improves principle alignment from 0.552 to 0.673 (Fig. 4). This matters because each Stage-II rollout costs a 42-minute P&R run on average. Incomplete placements are rejected before P&R and only rank last in their group, carrying no PPA signal; overlapping placements are legalized at a displacement penalty, so their measured PPA partly reflects the legalizer rather than the policy. Even starting from Stage-I, 59 of 688 logged Stage-II candidates were incomplete (App. A.5). On PlaceReasoner-Bench, Stage-I also routes 16/16 tasks versus 15/16 for SFT (Table 3).

W2/Q3:

We agree the original evidence mainly supports within-domain transfer, and we have revised the wording to say so (App. A.5 and A.7). We have added two pieces of evidence:

(1) Leave-family-out (App. A.7, Table 8). We retrained Stage-II from the same Stage-I checkpoint after removing every Ariane- and BlackParrot-derived design, then evaluated on those families (WNS / TNS, ns):

TaskStage-IStage-II (full)Stage-II (family held out)
ariane133 1:1-0.06 / -0.84-0.01 / -0.21-0.03 / -0.31
ariane133 2:1-0.12 / -11.33-0.06 / -2.52-0.06 / -2.66
ariane81 1:1-0.03 / -0.22-0.01 / -0.050.00 / 0.00
ariane81 2:1-0.16 / -83.17-0.02 / -0.42-0.07 / -0.82
bp_be 1:1-0.19 / -18.49-0.16 / -14.38-0.16 / -15.78
bp_be 2:1-0.33 / -32.06-0.23 / -21.84-0.24 / -22.15
bp_fe 1:1-0.04 / -0.09-0.04 / -0.14-0.03 / -0.13
bp_fe 2:1-0.17 / -7.70-0.07 / -0.28-0.07 / -0.29

With the families held out, Stage-II still improves WNS over Stage-I on all 8 tasks and TNS on 7/8, and stays within 0.05 ns WNS of the full model on every task. This indicates that the learned floorplanning decisions (placement style, halo, corridor, spacing) transfer across families rather than being memorized.

(2) Unseen design classes. vga_lcd, ethernet, and VeriGPU have no counterpart in training. On these 6 tasks, Stage-II improves TNS over Stage-I on 5 (e.g., VeriGPU 2:1: −223.85 → −13.92 ns; vga_lcd 2:1: −50.25 → −27.23 ns).

AuthorsOpenReview ↗

Response (2/2): Methodological novelty and timing vs. routability

W3: Methodological novelty.

We thank the reviewer for highlighting the need to characterize our contribution more precisely. We agree that SFT with LoRA and GRPO/DAPO-style RL are established methods. Our contribution lies in the problem formulation and training framework for learning open-weight macro-placement policies from physical PPA feedback, in a setting where access to real circuit data is limited and each P&R evaluation is expensive.

The framework addresses these constraints through three coupled design choices:

  • Learning placement competence without costly physical evaluations. A circuit-agnostic synthetic environment and verifier provide supervision for basic placement competence without proprietary circuits or P&R runs, preparing the policy for subsequent refinement using physical feedback.
  • Optimizing placement and flow decisions jointly. The Stage-II action space includes both macro coordinates and flow-level levers, allowing the policy to learn a floorplanning strategy whose downstream consequences are evaluated through the physical design flow.
  • Addressing exploitable verification failures. The reward formulation accounts for concrete failure modes in this setting. For example, the graded containment reward closes the overhang exploit permitted by binary containment (Sec. 3.3, Fig. 3).

The synthetic-to-physical progression connects these elements: inexpensive, verifiable supervision establishes basic placement competence before costly physical feedback guides the optimization of placement and flow decisions. We agree that evidence from macro placement does not establish a general learning principle beyond EDA. We have revised the contribution summary and conclusion to state the supported scope explicitly: a training framework for macro placement that connects synthetic supervision with optimization using physical feedback.

Q4: Timing vs. routability.

Stage-II improves timing and sharply reduces total DRVs, at the cost of fewer fully DRV-clean layouts. The full picture: Stage-II produces fewer DRC-clean layouts (8 vs. 11 of 16), but far fewer violations overall: 29 vs. 446 total DRVs across the 16 tasks, and a worst case of 8 vs. 393 (ethernet 2:1 under Stage-I). Prioritizing timing in the reward (weight 3.0 vs. 2.0 for DRC) is deliberate: small residual DRV counts are usually cleared by incremental routing repair, whereas slack is largely fixed by the macro placement. The revision now reports both the DRC-clean count and the total DRV count, and discusses rebalancing the DRC weight (Sec. 4.3).

We hope these responses address your concerns, and we are happy to answer any further questions during the discussion period.

Thank you for the response and additional experiments. I appreciate the new leave-family-out evaluation and the revised DRC/DRV reporting, which partially address my generalization and routability concerns. However, the response does not fully resolve my central concerns because the statistical evidence for the frontier-level comparison remains weak and the same-budget experiment needed to justify the progressive training curriculum is missing, while the methodological novelty concern is mainly addressed by reframing rather than by substantive advance. I therefore maintain my score.

Thank you for your reply. For the record, we briefly summarize the evidence on the two remaining points.

1. Statistical evidence for the Opus 4.8 comparison. With n=3, per-task CIs are wide, but a model-level comparison is better tested by pairing the two models across tasks. Using the per-task means already reported (App. A.6):

  • WNS: PR-Pro-II is better on 14/16 tasks (1 tie, 1 loss); Wilcoxon signed-rank p ≈ 0.001.
  • TNS: PR-Pro-II is better on 14/16 tasks; Wilcoxon signed-rank p ≈ 0.003.
  • At the design level (averaging the two aspect ratios, which are not independent): PR-Pro-II is better on 8/8 designs for both WNS and TNS (exact p ≈ 0.008).

The timing comparison therefore does not rest on particular sampled placements. Our headline claim remains scoped as comparable timing with a remaining routability gap.

2. Why Stage I RL precedes Stage II RL. Each Stage-II rollout costs about 42 minutes of P&R, and rollouts with geometric failures waste it: incomplete placements carry no PPA signal, and overlapping placements are measured only after legalization. Stage I removes these failures with no P&R cost. Compared with SFT, it raises overlap-free placement from 78.1% to 90.9% and containment from 84.3% to 92.9%, and it routes 16/16 tasks versus 15/16. Even from Stage I, 59 of 688 Stage-II candidates were incomplete, so an SFT start would waste more of the budget. A matched-budget SFT → Stage II run (~480 P&R-hours) could not finish within the discussion period.

3. Novelty. As stated in our response, our contribution is a training framework for macro placement that combines synthetic supervision with physical-feedback optimization. The revised paper states this scope explicitly.

We thank the reviewer again for the constructive discussion.

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

The paper specializes an 8B vision-language model for macro placement through synthetic supervised fine-tuning, geometric reinforcement learning, and reinforcement learning with measured post-route feedback. Stage II jointly selects macro coordinates/orientations and floorplanning levers such as halos, corridors, and spacing. Evaluation uses 288 synthetic layouts and 16 physical-design tasks from eight designs at two aspect ratios. The paper reports improved geometric validity and timing, and compares the specialized model with a proprietary model within the same harness.

Critical strengths
  1. The progression from inexpensive geometric supervision to expensive measured PPA is a useful system design for domain specialization. A compact model that can run locally is practically relevant when external inference is undesirable, and the synthetic environment provides a reusable source of training supervision.

  2. The evaluation measures routed outcomes rather than only placement proxies. Stage II improves TNS over Stage I on 13 of 16 tasks (Table 3, p. 9; full tables on pp. 18–19). Summing the reported Table 3 DRV entries also gives a reduction from 446 to 29 total violations, despite fewer entirely violation-free tasks. The geometric results and actual reward-aggregation, curriculum, group-size, and scorer ablations (pp. 15–17) provide further evidence about the pipeline.

Critical weaknesses
  1. The frontier-level claim needs a more precise timing–violation comparison and uncertainty estimate. Table 3 (p. 9) gives 8/16 zero-DRV tasks for Stage II, 11/16 for Stage I, and 13/16 for PR-Beta. However, Stage II also reduces total DRV from 446 to 29 relative to Stage I; it is therefore inaccurate to characterize routability as simply worsening. PR-Beta has 4 total DRV in the same table. The authors already discuss the clean-task trade-off in Section 4.3 (p. 8), and their reward explicitly includes violations (Section 3.3). My concern is how to interpret broad placement competitiveness across these trade-offs, not an alleged omission of routability. The supplied tables do not report repeated-run uncertainty. The common harness/EDA flow and default temperature of 0.1 are specified (pp. 7, 13), but the number of evaluation candidates, selection procedure, and repeated generations behind each reported result remain insufficiently explicit. A comparison under the stated multiobjective criterion, with these evaluation details, would make the headline claim more convincing.

  2. Some configuration reporting and contribution attribution need clarification. Appendix A.5 explicitly states that no evaluation task is shared with training, acknowledges related Ariane/BlackParrot families, and motivates within-class specialization (p. 14); I do not regard this disclosed overlap as established leakage or a missing discussion of scope. Likewise, Appendix A.6 already tests the Stage-I reward curriculum and other choices (pp. 15–17), so an assertion that training order has not been ablated would be too broad. The narrower attribution limit is that the staged pipeline does not isolate the incremental value of reasoning text over the same coordinate supervision (Section 3.2, p. 5). This can be handled by qualifying mechanism claims, without requiring a new research direction. For reproducibility, the training set is described as 45 tasks, but the rollout loop says 42 (pp. 14–15). Section 4.4 also advertises no-alignment and direct-frontier-judge arms (pp. 8–9), whereas Table 7 reports base-versus-learned scorers (p. 17). The base scorer itself is already described in Section 3.3 (p. 6); the issue is reconciling the reported experimental arms, not claiming that no scorer is specified.

Questions for authors
  1. Given the shared harness and stated default temperature, how many evaluation generations underlie each Table 3 entry, and how are candidates selected? How should the timing comparison be interpreted jointly with the reported zero-DRV counts and aggregate violations under the chosen multiobjective criterion?

  2. Does Stage II cycle through 42 or 45 tasks? Were the no-alignment and direct-frontier-judge experiments advertised in Section 4.4 conducted, and where are their results relative to the base-versus-learned scorer comparison in Table 7?

  3. Beyond the existing reward-curriculum ablation, is there evidence separating the contribution of reasoning-text supervision from the overall post-training pipeline? If not, can the interpretation be limited to the demonstrated pipeline rather than attributing its improvement specifically to reasoning traces?

1 comment
AuthorsOpenReview ↗

Clarified evaluation protocol, corrected Sec. 4.4, and new reasoning-text ablation (App. A.9)

Thank you for the careful reading and for the precise, actionable suggestions. We address each point below.

W1/Q1: Evaluation protocol and how to read the timing-routability results.

Evaluation protocol. For each task, the policy generates 5 candidates (5 independent trajectories) at temperature 0.1. Every candidate runs through the identical P&R flow. A candidate counts as a failure if it does not finish detailed routing or has more than 200 DRVs. Among valid candidates, we select based on the priority shown in Fig. 2: timing > DRV > power = area. Note that the same protocol applies to PR-Beta and all PR-Pro variants. We have added these details to Section 4.1 and Appendix A.4.

Interpretation. Relative to Stage I, Stage II improves timing on most tasks (TNS on 13/16) and reduces total DRVs from 446 to 29, but it has fewer DRV-clean tasks (8 vs. 11). The remaining violations are small in absolute terms. No Stage-II result has more than 8 DRVs, and three cases account for 21 of the 29. Residual counts at this scale are typically resolved by incremental routing repair in standard flows, so the practical concern is keeping DRVs low rather than eliminating them at placement. Relative to PR-Beta, Stage II is comparable on timing (7 vs. 7 WNS wins, 7 vs. 9 TNS wins) but less clean on routability (29 vs. 4 total DRVs; 8 vs. 13 clean tasks). Under our timing-first criterion, the accurate claim is therefore comparable timing with a remaining routability gap, not broad frontier-level placement quality.

W2/Q2: (i) Does Stage II cycle through 42 or 45 tasks? Stage II cycles through 45 tasks; "42" was a typo and has been corrected. Thank you for catching this inconsistency.

(ii) Were the no-alignment and direct-frontier-judge experiments advertised in Section 4.4 conducted? We did not run the no-alignment or in-loop frontier-judge variants, and we have revised Sec. 4.4 to list only the experiment we ran. We deliberately did not put a frontier judge in the RL loop, as that would reintroduce the external-inference dependency our method is designed to remove. The experiment we ran (App. A.9, Table 11) compares two in-loop scorers: an 8B scorer distilled from frontier-judge scores, and the zero-shot base model. We evaluate both resulting policies under the same frontier judge. Geometric metrics are unchanged (within one point), but the learned-scorer policy reaches lower rubric alignment (0.61 vs. 0.67), which we attribute to the policy exploiting cases where the scorer disagrees with the judge it was distilled from. The final pipeline therefore uses the base scorer.

Q3: Contribution of reasoning-text supervision.

Reasoning-text supervision contributes substantially: it raises design-principle alignment by +0.121 and overlap-free placement by +9.7 points over an otherwise identical answer-only model, giving Stage-I RL a stronger starting point.

To isolate this effect, we trained an answer-only SFT baseline on the same 10K placements with the reasoning traces removed, keeping the data, LoRA settings, and training budget identical. Results on PlaceReasoner-Layout-Bench (288 tasks):

ModelAlignCompleteIn-coreOverlap-free
Qwen3-VL-8B base0.340.2490.2290.125
SFT, answer only0.4311.0000.8680.684
SFT, with reasoning0.5520.9780.8430.781
Stage-I RL (from SFT with reasoning)0.6730.980.9290.909

Adding reasoning traces raises alignment from 0.431 to 0.552 and overlap-free placement from 0.684 to 0.781, with only slight decreases in containment (−2.5 pts) and completeness (−2.2 pts). Stage-I RL then raises containment to 0.929, overlap-free placement to 0.909, and alignment to 0.673. Reasoning supervision is therefore a key contributor to design-principle alignment, while most of the geometric legality comes from Stage-I RL. Thus, the interpretation cannot be limited to the demonstrated pipeline. Reasoning traces also contribute a lot to the improvement. We have added this ablation to the revised paper ([Appendix A.9]), all revisions are highlighted in blue.

We hope these responses address your concerns, and we are happy to answer any further questions during the discussion period.

Comments on the paper

1 comment

Dear Reviewers,

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

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.