acceptodds
Under review as a conference paper at ICLR 2027

From Critical Paths to Critical Cells: Learning Fine-Grained Timing Attribution for Chip Design

Abstract

Identifying which components of a system to improve is difficult when supervision is available only for aggregated outcomes. Timing optimization in chip design presents this challenge: static timing analysis (STA) identifies timing-critical paths, while placement refinement is made on cells whose contributions to a path's timing degradation may differ. Identifying a critical path does not by itself reveal the cells whose optimization would most improve timing. We study **timing-critical cell identification**, the task of learning cell-level timing guidance from path-level observations without cell-level annotations. We introduce CritiCell, a multiple instance learning framework that represents timing paths as bags of cells and learns cell scores from their aggregate timing behavior. We also utilize these scores to guide both constraint-driven and objective-driven placement in the ICCAD2015 benchmark suite. Compared with coarse path-level guidance, our fine-grained CritiCell improves worst negative slack (WNS) and total negative slack (TNS) by 2.87% and 10.74% in constraint-driven placement, and by 4.08% and 10.56% in objective-driven placement. These results show that coarse supervision can produce fine-grained guidance useful for downstream optimization tasks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Timing-Critical Cells, Weakly Supervised Learning, Multiple Instance Learning, Electronic Design Automation, Timing Optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vNdjOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CritiCell, a multiple instance learning framework that infers cell-level timing criticality from path-level labels and guides two placement optimization workflows.

Critical strengths
  1. Addresses a practically relevant problem without requiring cell-level annotations.

  2. Two placement applications, feature ablations, and heuristic comparisons support its effectiveness.

Critical weaknesses
  1. The max-pooling assumption may not adequately capture the cumulative nature of path delays. Placement improvements alone do not establish accurate cell-level attribution.

  2. Main results aggregate training, validation, and unseen circuits while using circuit-specific hyperparameters, limiting the evidence for generalization.

Questions for authors
  1. Can controlled local-movement experiments with equal optimization budgets verify that highly scored cells offer greater timing improvement potential?

  2. How does the method perform across multiple random seeds on unseen circuits with fixed hyperparameters?

5 comments
AuthorsOpenReview ↗

Response to Reviewer vNdj — Part 1

Dear reviewer vNdj,

Thank you for your insightful comments. We sincerely hope our rebuttal could adequately address your concerns. If so, we would deeply appreciate it if you could consider raising your score. If not, please let us know your further concerns, and we will continue actively responding to your comments.

Question 1

Thank you for suggesting this test. For each of the three unseen circuits, we sampled 16 timing-violating paths from a legalized DREAMPlace placement. On each path, we selected the cell with the highest CritiCell score, the cell with the highest local-delay score, and a uniformly random cell.

Each selected cell received the same search budget of eight predefined displacement candidates within ten row heights, plus a no-op. We applied each legal move independently from the same baseline placement, recomputed placement parasitics, and reran STA. For each cell, we recorded the largest delay reduction on the same fixed data path among the tested moves.

Mean best path-delay reduction across 16 paths per circuit (ps)

CircuitCritiCell (Ours)Local-delayRandom
superblue109.7315.8144.633
superblue1614.85810.2040.931
superblue187.2923.9172.845

Under the same local search budget, cells selected by CritiCell yielded larger mean path-delay reductions than either baseline on all three unseen circuits, showing that cells with high CritiCell scores offer greater timing improvement potential.

We have added this equal-budget local-movement experiment to Section 4.3 of the revised paper.

Question 2

Thank you for asking about sensitivity to random seeds. We evaluated Enhanced Efficient-TDP with three placement seeds on each of the three circuits unseen by the trained model. Since timing conditions vary across circuits, we used a shared empirical rule to set the two attraction multipliers based on the initial STA feedback for each run, then kept those values fixed for the remainder of placement.

Let NN and HH denote the counts of repeated non-critical and critical pin-pair occurrences, respectively, divided by the number of movable cells.

αnon=clip⁡(N−0.4,0,2),αcrit=clip⁡(0.04/H,1,2),\alpha_{\mathrm{non}}=\operatorname{clip}(N-0.4,0,2), \qquad \alpha_{\mathrm{crit}}=\operatorname{clip}(0.04/H,1,2),

with αcrit=1\alpha_{\mathrm{crit}}=1 when H=0H=0.

The table reports WNS / TNS improvements for Enhanced Efficient-TDP over the original Efficient-TDP on the same circuit and placement seed.

CircuitSeed 4001Seed 4002Seed 4003
superblue10−1.29% / −5.56%+3.27% / +1.15%+2.14% / −1.45%
superblue16+1.75% / +7.60%+4.18% / +1.51%+28.56% / +16.97%
superblue18+6.67% / +10.00%+3.93% / +11.71%+8.20% / +15.98%

Averaged equally across circuits, WNS and TNS improved by 6.38% and 6.43%, respectively, with both improving in 7 of 9 paired comparisons. These results demonstrate the robustness of our method.

We have added the shared parameter rule and the multi-seed results to Appendix C.9 of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer vNdj — Part 2

Weakness 1

Thank you for raising these points. STA accounts for cumulative delays along each path when computing slack. In CritiCell, max pooling aggregates learned cell scores to predict the binary path label (Eq. 4). This lets the model learn from path-level supervision under the assumption that a timing-critical path contains at least one bottleneck cell. Our training filter excludes long violating paths to reduce cases where delay accumulated across many stages dominates the path label (Section 3.3).

The controlled single-cell moves in our response to Question 1 provide direct cell-level evidence. Under the same local search budget, cells with high CritiCell scores yielded larger mean path-delay reductions than either baseline on all three unseen circuits.

Weakness 2

Thank you for raising this point. We report the mean per-circuit WNS / TNS gains over the timing-aware baselines in Tables 1–2, grouped by the model development split in Appendix B.3.

ComparisonTrainingValidationUnseen test
Ours (Clust.) vs. TritonPart+11.82% / +12.40%−6.52% / +7.11%+3.16% / +12.06%
Ours (TDP) vs. Efficient-TDP+5.46% / +8.79%−0.98% / +15.48%+6.08% / +9.05%

Both placement flows improve WNS and TNS on average on circuits held out from model development. To assess performance with a common parameter-selection procedure, our response to Question 2 reports a multi-seed evaluation in which a shared rule sets hyperparameters. In comparisons using the same placement seed for both methods, Enhanced Efficient-TDP improved WNS by 6.38% and TNS by 6.43% on average over the original Efficient-TDP.

We have added the results by data split to Appendix B.3 of the revised paper.

AuthorsOpenReview ↗

Request for Feedback on Our Responses

Dear Reviewer vNdj,

We would appreciate it if you could review our responses when you have a chance. We hope they address your concerns and clarify our work. If you have any further questions or concerns, please let us know. We would be happy to continue the discussion and provide additional clarification.

Thank you again for your time and thoughtful feedback.

Best regards,
The Authors

Reviewer_4v7fOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how to turn path-level timing feedback from static timing analysis (STA) into cell-level guidance for placement. It proposes CritiCell, a multiple-instance learning (MIL) model that treats each timing path as a bag of cells and predicts a criticality score for each cell in the context of that path. The model uses 16 geometric, timing, positional, and connectivity features, with binary path labels derived from STA as supervision. The training paths are filtered by slack and length to better fit the MIL assumption. The resulting scores guide clustering-based placement and modify the pin-pair attraction weights in Efficient-TDP. The reported timing improvements are promising, but the paper does not yet establish that the learned scores identify the cells whose placement changes would yield the largest timing gains.

Critical strengths
  1. Well-motivated problem. STA reports timing violations at the path level, whereas placement acts on individual cells. Distinguishing the cells that deserve attention could make timing-driven placement more selective and reduce unnecessary changes elsewhere in the design.

  2. Practical use of existing supervision. The MIL formulation avoids the need for manually labeled critical cells and makes use of timing reports already produced in a standard EDA flow. The model is lightweight, and its inputs are readily obtainable from placement and STA.

Critical weaknesses
  1. The need for MIL is not convincingly established. STA already exposes cell-arc delays, net delays, and pin-level timing information, yet the paper gives limited evidence that the learned scores are better than strong STA-based alternatives. More fundamentally, a cell that contributes substantially to path delay is not necessarily one whose placement can be changed to improve timing.

  2. Path coverage effects are not modeled explicitly by the predictor. A cell may have modest importance on each of many critical paths yet be highly relevant to overall TNS; a fixed per-path score threshold could miss such a cell. The downstream optimizers account for some path overlap through group merging or accumulated pin-pair weights, but this does not resolve the limitation in the learned scores.

  3. The optimization contribution is limited, and the objectives are only indirectly connected. Both applications rely heavily on existing placement methods, with critical-cell information mainly used to alter grouping preferences or attraction weights. The MIL loss trains for binary path classification, while the placement objective penalizes weighted pin distances; neither directly measures the timing benefit of moving a particular cell.

Questions for authors
  1. Strengthen the STA-based baselines. Compare against criticality measures based on cell/net delay, pin slack, path coverage, and slack-weighted path aggregation, not only local-delay ranking and a handcrafted composite score. Use comparable cell-selection budgets so that differences in the number of targeted cells do not confound the results.

  2. Test whether the scores predict actual placement sensitivity. Apply controlled, legal local placement changes to selected cells, rerun STA, and measure the resulting WNS and TNS changes. The correlation between these gains and the MIL scores would provide much more direct evidence for the paper's central claim.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 4v7f — Part 1

Dear reviewer 4v7f,

Thank you for your insightful comments. We sincerely hope our rebuttal could adequately address your concerns. If so, we would deeply appreciate it if you could consider raising your score. If not, please let us know your further concerns, and we will continue actively responding to your comments.

Question 1

Thank you for your suggestion. We evaluated six additional STA-based selectors in the Enhanced Efficient-TDP flow on the three circuits held out from model development. They cover the four requested categories:

  • Cell/net delay: cell-arc delay, the sum of the two adjacent net delays, and the maximum of those delays.
  • Pin slack: the magnitude of the worst negative slack among a cell's data pins.
  • Path coverage: the number of distinct timing-violating paths that contain the cell.
  • Slack-weighted path aggregation: the sum of negative slack magnitudes over those distinct paths.

At each timing update, we used the number of critical cells identified by CritiCell Enhanced Efficient-TDP in each STA-based baseline's current placement as that baseline's selection budget. Each baseline ranked the candidates by its own STA score, selected exactly that many, and used the same pin-pair weighting rule as CritiCell Enhanced Efficient-TDP.

The table reports WNS / TNS gains (%) over the Efficient-TDP values in Table 2. Ours is the published Table 2 result.

Selectorsuperblue10superblue16superblue18Mean
Cell-arc delay−3.65 / −0.25−0.16 / +16.43+10.35 / +14.34+2.18 / +10.18
Adjacent net delay (sum)−2.38 / −1.32+5.65 / +15.19+5.80 / +0.09+3.02 / +4.66
Adjacent net delay (max)−2.13 / +1.68+5.31 / −3.47+7.24 / +0.02+3.47 / −0.59
Pin slack−7.38 / +1.22+1.66 / +19.58+0.33 / −0.31−1.79 / +6.83
Path coverage−4.86 / −1.03−14.34 / −12.29+8.09 / +4.51−3.70 / −2.94
Slack-weighted path aggregation−2.04 / −0.45+8.29 / +18.71+6.88 / −0.06+4.38 / +6.07
Ours (TDP, Table 2)+0.48 / +0.43+7.07 / +16.07+10.69 / +10.64+6.08 / +9.05

In this comparison, Ours achieves the highest mean WNS gain and a higher mean TNS gain than five of the six STA baselines. Cell-arc delay yields a slightly higher mean TNS gain (10.18% vs. 9.05%) and a substantially lower mean WNS gain (2.18% vs. 6.08%). Ours shows more consistent gains across the three circuits, improving both metrics on all three. Cell-arc delay improves both only on superblue18. It worsens both WNS and TNS on superblue10 and worsens WNS on superblue16.

We have added the six STA-based baselines and the matched-budget comparison to Appendix C.8 of the revised paper.

Question 2

Thank you for suggesting a controlled test of placement sensitivity. CritiCell assigns each cell a score in the context of a specific timing path. We therefore expect a path-specific MIL score to have a limited correlation with design-level WNS and TNS changes. However, using CritiCell to identify critical cells across a design and guide placement yields substantial overall gains in both WNS and TNS (Tables 1–2).

To provide more direct evidence for our central claim, we tested whether cells prioritized by CritiCell yield larger path-delay reductions under controlled local moves. We sampled 16 timing-violating paths on each of the three unseen circuits. On each path, we selected the highest-scoring cell under CritiCell, the cell with the highest local-delay score, and a uniformly random cell. Starting from a legalized DREAMPlace placement, we generated legal moves for each cell using the same eight predefined displacements and included a no-op. We applied each move independently from the same baseline placement, recomputed placement parasitics, and reran STA. For each selected cell, we recorded the largest delay reduction along the same fixed data path among its tested moves.

Mean best path-delay reduction across 16 paths per circuit (ps)

CircuitCritiCell (Ours)Local-delayRandom
superblue109.7315.8144.633
superblue1614.85810.2040.931
superblue187.2923.9172.845

On all three circuits, moving the cells selected by CritiCell yielded larger mean path-delay reductions than moving the cells selected by either baseline. These controlled interventions provide direct evidence that CritiCell identifies cells with greater timing sensitivity.

We have added the controlled local-movement experiment and its results to Section 4.3 of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer 4v7f — Part 2

Weakness 1

Thank you for raising these points. MIL lets CritiCell learn path-specific cell scores from STA-derived path labels without cell-level annotations, combining delay features with placement geometry, fanout, and path position.

As shown in our response to Question 1, CritiCell compares favorably with strong STA-based selectors under comparable cell-selection budgets.

The controlled single-cell moves in our response to Question 2 show that cells selected by CritiCell offer greater potential for timing improvement.

Weakness 2

Thank you for raising this point. CritiCell identifies bottleneck cells within timing-critical paths. It adds cell-level priorities to path-level STA feedback so that the placer can focus more effort on those bottlenecks.

Cells with modest importance on individual paths but a substantial cumulative impact on TNS can still receive timing guidance based on the original STA reports, even if their CritiCell scores fall below the threshold on every path. For example, Enhanced Efficient-TDP applies a slack-weighted attraction update to every adjacent pin pair on the critical paths it considers, with smaller updates for pairs between non-critical cells. A cell with modest importance on each of many critical paths can therefore still be subject to substantial attraction through the accumulated weight updates across those paths (when αnon>0\alpha_{\mathrm{non}} > 0).

Enhanced Efficient-TDP improves WNS by 4.08% and TNS by 10.56% on average over the original Efficient-TDP (Table 2). Under comparable cell-selection budgets, CritiCell also achieved larger mean gains in both metrics than the path coverage and slack-weighted path aggregation selectors (our response to Question 1).

Weakness 3

Thank you for raising this point. Our optimization contribution is to use cell-level priorities learned from path-level supervision to guide timing optimization. The controlled experiment in our response to Question 2 directly measures the timing benefit of moving a selected cell.

Reviewer_4v7fOpenReview ↗

Any advantages comparison with previous works

Dear authors,

Thank you for your reply. However, I have several critical concerns about the motivation and novelty.

  • Timing-critical cell identification relies on the neighbouring cells. Previous works use GNN-based methods to identify them. However, the GNN-based models have limited scope to capture the influence of neighbouring cells. I am not clear why your proposed method can handle the limitation. In addition, critical path identification is a path-based problem. Based on the identified cells, how can you guarantee that you can accurately find the critical path?
  • Commercial tools such as PT and industrial tools such as OpenSTA and OpenTimer have good performance in terms of accuracy and runtime. I am not sure whether learning-based methods need to be proposed.
  • Similar techniques have been reported in previous works, such as [Ye et al., 2023b] and [Liao et al., 2022]. The proposed method just combines previous works incrementally. The authors are suggested to compare with [Liao et al., 2022].
Reviewer_K2GUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a weakly supervised framework for identifying timing-critical cells in chip design. It addresses the mismatch between path-level timing feedback from static timing analysis (STA) and cell-level decisions in placement optimization. Specifically, the method formulates each timing path as a bag of cells, uses a shared MLP to predict cell-level importance scores, and applies max pooling to obtain path-level predictions for training, without requiring cell-level annotations. The learned scores are then incorporated into clustering-based placement and an enhanced version of Efficient-TDP.

The main contribution is to demonstrate how path-level supervision can be used to derive fine-grained timing guidance for different placement optimization strategies. Experiments on eight ICCAD2015 benchmark circuits show improvements in worst negative slack (WNS) and total negative slack (TNS) over DREAMPlace, TritonPart-based placement, and the original Efficient-TDP. The authors further evaluate the method through feature ablations, path-level consistency tests, and qualitative case studies. The reported results suggest that the learned guidance can improve downstream timing optimization. The evaluation provides evidence of practical benefits, but the validity of the cell-level attribution remains difficult to establish. Since no cell-level ground-truth labels are available, the path-level validation only measures consistency with STA-derived path labels, rather than whether the selected cells are actually responsible for timing degradation.

Critical strengths

The paper proposes using multiple-instance learning to infer cell-level criticality from path-level timing labels and integrates the resulting scores into two timing-driven placement pipelines. The reported improvements in WNS and TNS suggest that the learned scores can provide useful guidance for downstream optimization.

Critical weaknesses

W1. Lack of direct validation of cell-level attribution. The model is trained with path-level labels, but the paper provides no direct evidence that its scores identify the cells responsible for timing degradation. Improvements in WNS and TNS demonstrate practical utility, but do not establish the accuracy of the proposed attribution.

W2. Inconsistent performance across circuits. The reported timing improvements are not consistent across all circuits, with some showing limited or negative gains. This raises concerns about the robustness of the proposed method and its ability to generalize across different designs. The paper does not sufficiently explain these variations or identify the conditions under which CritiCell is effective.

Questions for authors
  1. How do the authors verify that cells assigned high criticality scores are truly responsible for timing degradation? Improvements in WNS and TNS demonstrate downstream optimization benefits, but do not directly establish the correctness of cell-level attribution. What evidence supports the claim that CritiCell identifies the cells that should be prioritized for optimization?

  2. The reported results aggregate performance across training, validation, and unseen test circuits, potentially masking differences in generalization. Why does the proposed method fail to consistently outperform the baselines on the validation circuits? Can the authors report results separately for training, validation, and unseen test circuits to clarify whether the gains generalize beyond the training data?

  3. Appendix C.7 evaluates an alternative data split only with Enhanced Efficient-TDP. Why was the clustering-based placement experiment not repeated under the same split, and do the results support robustness across both placement methods?

3 comments
AuthorsOpenReview ↗

Response to Reviewer K2GU — Part 1

Dear reviewer K2GU,

Thank you for your insightful comments. We sincerely hope our rebuttal could adequately address your concerns. If so, we would deeply appreciate it if you could consider raising your score. If not, please let us know your further concerns, and we will continue actively responding to your comments.

Weakness 1 & Question 1

Thank you for your comments on cell-level attribution. To test cell prioritization directly, we measured how much path delay could be recovered by legally moving individual cells under a fixed search budget.

For each of the three circuits unseen during model development, we used a legalized DREAMPlace placement as the baseline and sampled 16 timing-violating paths from it. On each path, we selected the cell with the highest CritiCell score, the cell with the highest local-delay score, and a uniformly random cell. For each target, we used the same eight predefined displacement candidates to generate legal moves and included a no-op. We applied each move independently from the same baseline placement, recomputed placement parasitics, and reran STA. For each target, we recorded the largest delay reduction along the same fixed data path among its candidates.

Mean best delay reduction over 16 paths per circuit (ps, higher is better).

CircuitCritiCell (Ours)Local-delayRandom
superblue109.7315.8144.633
superblue1614.85810.2040.931
superblue187.2923.9172.845

Cells selected by CritiCell yielded larger mean delay reductions than those selected by either baseline on all three circuits. Because each intervention changed only the selected cell's location, the measured reductions quantify the delay recoverable by moving that cell within the tested candidate set. This supports using CritiCell scores to prioritize cells for local placement search.

We have added this controlled local-movement experiment to Section 4.3 of the revised paper.

Weakness 2

Thank you for highlighting the variation across designs. Differences in critical-path structure and circuit characteristics can affect the timing gains available from cell prioritization. CritiCell improves both WNS and TNS on six of eight circuits against Efficient-TDP (Table 2). The mean WNS/TNS gains are 4.08%/10.56% over Efficient-TDP and 2.87%/10.74% over TritonPart (Tables 1–2). CritiCell improves at least one timing metric in every reported per-circuit comparison with these two baselines, making its WNS/TNS outcomes Pareto-nondominated in those pairwise comparisons.

To examine why the gains can be modest, we analyzed the critical paths of superblue10, which shows small improvements in both WNS and TNS over Efficient-TDP.

For each design, we analyzed the first 1,000 timing-violating paths extracted from the placement generated by DREAMPlace. For each path, we measured its number of cell arcs and the share of total data-path delay accounted for by the largest local segment, defined as a cell arc and its two adjacent net arcs. Among the eight designs, superblue10 had the highest median arc count and the lowest median delay share for the largest local segment.

Held-out designMedian cell arcs per pathMedian delay share of the largest local segmentMean relative path-delay reduction with CritiCell
superblue103038.33%0.0492%
superblue165622.28%0.1794%
superblue186431.51%0.1087%

The last column comes from the single-cell movement experiment in W1 + Q1. For each of the 16 paths per design, we divided the best tested delay reduction for the cell selected by CritiCell by that path's baseline delay, then averaged across paths. The smaller relative reduction on superblue10 is consistent with the idea that moving a few cells can recover less of the total delay when it is spread across many stages. These observations suggest that cell prioritization may offer greater timing benefits on designs where a substantial share of critical-path delay is concentrated in a few placement-sensitive segments.

AuthorsOpenReview ↗

Response to Reviewer K2GU — Part 2

Question 2

Thank you for your question. Below are mean per-circuit gains over the timing-aware baselines in Tables 1–2, using the split in Appendix B.3.

ComparisonTraining WNS / TNSValidation WNS / TNSUnseen test WNS / TNS
Ours (Clust.) vs. TritonPart+11.82% / +12.40%−6.52% / +7.11%+3.16% / +12.06%
Ours (TDP) vs. Efficient-TDP+5.46% / +8.79%−0.98% / +15.48%+6.08% / +9.05%

We use the validation circuits for model selection. The timing gains come from applying the selected model to guide placement and can vary with circuit characteristics, as discussed in W2. In Table 2, both validation circuits show TNS gains and small WNS regressions.

Both comparisons show positive mean WNS and TNS gains on unseen circuits, supporting transfer beyond the training circuits under the reported settings.

We have added the results by data split to Appendix B.3 of the revised paper.

Question 3

Thank you for your question. Appendix C.7 evaluated the alternative split using Enhanced Efficient-TDP. We extended the evaluation to Clustering-Based Placement on all eight circuits under the same split.

As in Table 1, WNS is in units of 10310^3 ps and TNS is in units of 10510^5 ps. The baseline timings are reported in Table 1.

DesignOurs (Clust.) WNSOurs (Clust.) TNS
superblue1−19.42−253.06
superblue3−28.31−130.23
superblue4−19.27−205.12
superblue5−45.71−196.33
superblue7−13.71−214.62
superblue10−26.22−912.90
superblue16−16.45−229.37
superblue18−12.27−72.59
Mean gain vs. DREAMPlace+15.08%+12.01%
Mean gain vs. TritonPart+4.77%+14.13%

Together with the 4.26% WNS and 8.22% TNS gains over Efficient-TDP in Table 13, these results show timing gains in both placement flows under the alternative data split. This highlights CritiCell's robustness to changes in the data split.

We have added the clustering-based placement results under the alternative data split to Appendix C.7 of the revised paper (Table 14).

AuthorsOpenReview ↗

Request for Feedback on Our Responses

Dear Reviewer K2GU,

We would appreciate it if you could review our responses when you have a chance. We hope they address your concerns and clarify our work. If you find that your concerns have been sufficiently addressed, we would be grateful if you could consider raising your score. If any questions or concerns remain, please let us know. We would be happy to continue the discussion and provide additional clarification.

Thank you again for your time and thoughtful feedback.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.