acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Neural Combinatorial Optimization for Job Scheduling Problems via Transition-Aware Dual-Scale Solver

Abstract

Neural scheduling policies face a representation–computation trade-off: compact candidate features can leave important action-induced changes implicit, whereas encoding complete successor states repeatedly processes largely unchanged information. Meanwhile, dependencies accumulate across decisions, while competition must be resolved among a changing set of feasible candidates, requiring both persistent context and detailed local comparison. To address these challenges, we propose the Transition-Aware Dual-Scale Solver (TADS). TADS combines a transition-complete action representation with a dual-scale encoder. For each feasible action, a task-specific compiler records the induced state changes as sparse typed edits. These edits reconstruct the canonical successor from the current state, while an order-invariant encoder maps each edit set to a transition embedding. The dual-scale encoder compares candidates through global–local reasoning. Global recurrent memory integrates the residual-state summary with the previously executed transition embedding to maintain context across decisions. Conditioned on this context, relation-aware local attention compares candidate transition embeddings and their scheduling relationships to guide action selection. The selected transition embedding feeds the next memory update, allowing subsequent decisions to account for the committed change. Experiments with independently trained policies across six scheduling formulations demonstrate competitive solution quality on synthetic test sets and public benchmarks.

Primary area:
optimization
Keywords:
Neural Combinatorial Optimization, Job Scheduling, Reinforcement Learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_WaL2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes the Transition-Aware Dual-Scale Solver (TADS) for constructive scheduling. Each feasible action is represented as a set of typed edits that, when applied to the current state, reconstruct its canonical successor. These edits are encoded into transition embeddings, and candidate actions are evaluated using global recurrent memory and local attention over candidates. The selected action’s embedding is fed into the next memory update. Experiments cover six scheduling formulations, including synthetic instances and public benchmarks, with component ablations and efficiency analyses. The reported results support the effectiveness of TADS.

Critical strengths
  1. TADS represents candidate actions through the state changes they induce and reuses the selected transition embedding to update memory. This connects action evaluation with subsequent decisions, allowing the model to incorporate executed action effects into its scheduling context.

  2. The dual-scale encoder combines global recurrent memory with local attention over candidate actions, enabling decisions to account for both historical context and current competition among candidates. Experiments across six scheduling formulations, including synthetic instances and public benchmarks, together with component ablations, provide empirical support for this design.

Critical weaknesses
  1. Transition completeness follows from recording all changed fields and reconstructing derived information under the stated assumptions. Although the representation comparisons support the effectiveness of edit encoding, they do not isolate completeness itself as the source of improvement. This limits the theoretical contribution and leaves its connection to scheduling performance unclear.

  2. The main text introduces many symbols and equations for relatively straightforward operations, such as recording changes and pooling features. This increases the reading cost and makes it difficult to identify the substantive novelty beyond combining transition features, recurrent memory, and candidate attention.

Questions for authors
  1. What evidence distinguishes the benefit of transition completeness from the benefit of explicit change encoding? The representation comparisons show performance differences, but full-successor representations also contain complete successor information.

  2. What is the benefit of feeding the selected transition embedding into memory, beyond updating memory from the current state summary?

3 comments
AuthorsOpenReview ↗

Response to Reviewer WaL2: Completeness, Feedback, and Contribution

Dear Reviewer,

Thank you for your insightful comments on completeness and our contribution. We distinguish theoretical guarantees from empirical gains. All revisions are highlighted in blue in the re-uploaded revised PDF.

W1: Theoretical role and link to performance. Section 4.1 and Appendix C.1 establish exact reconstruction from the current state and symbolic edits, before learned pooling. This gives task-specific compilers a common, testable contract and permits unchanged state to be shared across candidates. Table S13 (Appendix E.3) reuses Table 5's zero-feedback transition-complete (TC) and full-successor (FS) checkpoints on identical flexible job-shop scheduling problem (FJSP) instances with the same backbone and greedy decoding. TC achieves 0.12–0.14 percentage points (pp) lower mean Gap, 1.23–1.30× speedup, and 17%–25% lower peak GPU memory allocation. Latency includes feature preparation, compilation or successor construction, inference, selection, and execution. The theory supports an exact sparse interface; its scheduling value is demonstrated through joint quality–cost gains.

Metric10×520×1040×10
FSTCFSTCFSTC
Obj. ↓106.28106.14197.06196.82368.92368.45
Gap (%) ↓10.3410.200.550.430.810.68
Latency (ms) ↓5404402350182048003700
Peak GPU memory (MiB) ↓5344786010478
Speedup (×) ↑1.23×1.29×1.30×
Memory reduction (%) ↑17.023.125.0

Q1: Completeness versus explicit change encoding. TC and Primary are both symbolically complete: each retains all primary writes and reconstructs derived information deterministically; TC additionally exposes derived events. Their contrast preserves reconstructible successor information while varying its explicit exposure. Timing keeps timing edits only; FS encodes complete successors. All four preserve the same simulator rules; Table 5 fixes zero feedback, backbone, training, and test instances. At 40×10, Primary−TC is 0.23 pp (derived events), Timing−Primary 0.32 pp (field coverage), and FS−TC 0.13 pp (change encoding). The same ordering holds at smaller sizes. These controls support explicit change encoding as the empirical source of gains.

Instance sizeMetricTCPrimaryTimingFS
10×5Gap (%) ↓10.2010.4210.6910.34
ΔGap (pp)—+0.22+0.27+0.14
20×10Gap (%) ↓0.430.620.900.55
ΔGap (pp)—+0.19+0.28+0.12
40×10Gap (%) ↓0.680.911.230.81
ΔGap (pp)—+0.23+0.32+0.13

Q2: Feedback beyond state-summary updates. Memory receives a compressed summary of the Markov state; transition feedback supplies the selected action's committed change directly (Section 4.2). Zero still updates memory from the current state summary. We retain the TC candidate representation and recurrent/local backbone, changing only feedback: selected transition embedding (Transition), zero vector (Zero), or pre-decision action embedding (Action). Each variant is trained independently with five seeds under its feedback rule in both training and evaluation. Only the selected candidate supplies feedback. At 40×10, Transition's mean Gap is 0.39%, versus 0.68% for Zero and 0.53% for Action: gains of 0.29 and 0.14 pp. Action also improves on Zero by 0.15 pp. Thus, Zero tests value beyond state-summary updates; Action tests consequence encoding against a generic feedback channel. Table 4 reports means and standard deviations (SD), with Transition lowest across all six tasks.

Task / test sizeTransition ↓Zero ↓Action ↓
SMSP 5001.18 ± 0.101.60 ± 0.151.39 ± 0.12
UPMSP 100×67.48 ± 0.188.04 ± 0.257.74 ± 0.21
PFSP 100×5−0.16 ± 0.050.06 ± 0.07−0.04 ± 0.06
FFSP 100×122.27 ± 0.162.43 ± 0.202.36 ± 0.18
JSSP 100×202.22 ± 0.122.86 ± 0.172.50 ± 0.15
FJSP 40×100.39 ± 0.050.68 ± 0.080.53 ± 0.06

W2: Revised exposition and substantive contribution. To address your concern about notation obscuring novelty, we reorganized and simplified the Methods around compile–compare–execute–reuse. Section 4.1 retains two equations instead of five: symbolic reconstruction and transition embedding. Encoding/pooling details and reconstruction are in Appendices B.3.1 and C.1. The contribution is a sparse transition interface linking candidate evaluation to subsequent memory updates: it shares unchanged state and reuses the selected pre-attention embedding without re-encoding. Pooling, recurrence, and attention implement this flow. These changes reduce main-text formalism and foreground the substantive interface and reuse mechanism.

We sincerely appreciate your careful review, which helped clarify our contribution, simplify the presentation, and distinguish reconstruction from policy gains. If these revisions resolve your concerns, we would be grateful if you would consider revisiting your overall assessment. We welcome questions about the revised claims or results.

Thanks for your detailed response and additional experiments. Through my deep thinking, I'll maintain my score.

Dear Reviewer,

Thank you for taking the time to reconsider our revised manuscript and additional experiments. We sincerely appreciate your careful assessment and respect your perspective.

We would like to emphasize that we have directly addressed the central concerns raised in your original review. Specifically, we have clarified the theoretical role of symbolic completeness, introduced controlled comparisons to distinguish its role from the empirical benefits of explicit change encoding, and independently evaluated the contribution of executed-transition feedback beyond the current state summary. We have also substantially simplified the Methods to make the substantive contribution clearer.

Beyond the empirical improvements, we believe TADS offers a meaningful contribution to neural combinatorial optimization and scheduling by introducing a transition-centered decision framework that explicitly models candidate action consequences and reuses executed-transition information to guide subsequent decisions. This provides a reusable perspective on sequential decision-making in scheduling, where each action affects future resource availability and feasible choices.

We sincerely hope that these targeted revisions and additional evidence have satisfactorily addressed your original concerns and demonstrated the value of our contribution. If so, we would be truly grateful if you could reconsider your overall rating in light of these improvements.

Thank you again for your time and thoughtful consideration.

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

The paper proposes TADS, a neural constructive solver for job scheduling that explicitly represents the consequences of candidate actions. Task-specific compilers produce sparse typed edits describing changes to resource availability, job progress, and feasible choices. Together with the current canonical state, these symbolic edits reconstruct the successor state; learned embeddings of the edits support action scoring. A dual-scale encoder combines global recurrent memory with relation-aware attention over feasible candidates, and the selected transition embedding feeds the next memory update. Separate policies are trained with recurrent PPO for six scheduling formulations. Evaluation covers synthetic instances, generalization to larger unseen sizes, and public benchmarks without benchmark-specific fine-tuning. Component ablations, alternative candidate representations, training-seed analyses, and runtime profiling examine the sources of the reported improvements. The results indicate competitive solution quality and favorable computational costs under controlled comparisons.

Critical strengths
  1. The representation and architecture address a clear modeling problem. Sparse transition edits expose action consequences while sharing unchanged state information across candidates. Combining candidate comparison with recurrent context, and reusing the selected transition embedding for subsequent decisions, provides a coherent design. Comparisons against original action features, selected consequence features, and full successor representations support the value of explicitly organizing inputs around state changes.

  2. The experimental analysis is broad and reasonably thorough. Evaluation spans six scheduling formulations, five training seeds, larger unseen sizes, and established public benchmarks. Controlled component ablations and separate compilation, network, latency, and memory measurements provide useful evidence for the design choices. For example, Table S12 reports that on JSSP 100 × 20, TADS reduces complete-schedule latency from 35.1 to 24.5 seconds and peak memory from 245.8 to 58.7 MiB relative to full-successor representation with full-residual attention.

Critical weaknesses
  1. Comparisons with external baselines are not fully controlled. Literature baselines retain their original hardware and decoding budgets, and some synthetic comparisons use newly generated instances rather than identical test sets. Several public-benchmark improvements are also small, such as 0.14% versus 0.15% Gap on PFSP 100 × 5. These results establish competitiveness more clearly than superiority under a common computational budget.

  2. The independent contribution of executed-transition feedback is not fully isolated. Removing the transition representation changes both candidate scoring and feedback, while removing global memory eliminates the entire recurrent component. Although these ablations support the overall architecture, they do not separate the benefit of feeding back the selected transition embedding from that of better candidate representations or recurrent context alone.

Questions for authors
  1. Can you compare against strong baselines such as MP-ASIL and ReSched on identical instances and hardware, using matched time budgets to assess the solution-quality versus runtime trade-off?
  2. With the candidate representation and recurrent architecture held fixed, how does replacing the selected-transition feedback with a zero vector or a conventional action embedding affect performance?
3 comments
AuthorsOpenReview ↗

Response to Reviewer vzRk: Matched Comparisons and Feedback Ablation

Dear Reviewer,

Thank you for your constructive feedback on baseline fairness and transition feedback. We have conducted matched-budget baseline comparisons and an independent executed-transition feedback ablation. All revisions are highlighted in blue in the re-uploaded revised PDF.

W1 / Q1: Comparisons on identical instances and hardware. We evaluate our Transition-Aware Dual-Scale Solver (TADS) against the requested ReSched and MP-ASIL, testing ReSched's REINFORCE and proximal policy optimization (PPO) variants and adding DANIEL, on identical flexible job-shop scheduling problem (FJSP) instances and hardware with matched wall-clock inference budgets. All methods share the same instance-specific reference objectives. Each method uses its decoding strategy without additional local search; only complete feasible schedules obtained within the time limit enter best-so-far evaluation (Appendix D.6). Timing includes environment initialization, complete decoding, and best-schedule selection, with CUDA synchronized at both boundaries. Figure 6 reports the quality–time curves for 10×5, 20×10, and 40×10, using the 20-GPU-hour checkpoints from Table S19. At 100 seconds, TADS obtains Gaps of 4.04%, -2.10%, and 0.12%, versus 4.62%, -1.78%, and 0.32% for MP-ASIL. Its curve is lowest over the displayed time ranges. We also evaluate public FJSP benchmarks with a 100-second budget (Table S17, Appendix E.6). Public evaluation covers 10 Brandimarte instances and 40 instances in each Hurink group; none is used for training or checkpoint selection. TADS has the lowest mean Gap in all four benchmark groups. The mean Gap reductions over MP-ASIL are 1.18, 0.47, 0.42, and 0.06 percentage points (pp). Matched comparisons concern FJSP; PFSP/JSSP retain source protocols. To control training resources as well, all five methods were independently trained at 10×5 with 10/20 GPU-hours per run, sharing the training distribution and validation instances. Baselines retain their architectures and optimization procedures; GPU-hours are matched rather than update counts. Checkpoints were selected by validation performance within each budget, then evaluated without test-specific fine-tuning under 100-second inference budgets (Table S19, Appendix E.7). TADS has the lowest mean Gap in all six settings, improving over MP-ASIL by 0.20–0.58 pp. Gap (%) is reference-relative; lower is better.

MethodBrandimarteHurink edataHurink rdataHurink vdata
DANIEL9.25%8.90%4.85%0.65%
ReSched-REINFORCE6.55%8.02%4.98%0.77%
ReSched-PPO6.30%7.73%4.77%0.69%
MP-ASIL7.10%7.05%4.54%0.48%
TADS (Ours)5.92%6.58%4.12%0.42%
Method10 GPU-hours20 GPU-hours
10×520×1040×1010×520×1040×10
DANIEL6.90%−0.55%1.35%5.65%−1.10%1.07%
ReSched-REINFORCE6.40%−0.85%0.95%5.18%−1.38%0.69%
ReSched-PPO6.18%−1.02%0.78%4.92%−1.55%0.51%
MP-ASIL6.05%−1.20%0.60%4.62%−1.78%0.32%
TADS (Ours)5.65%−1.52%0.37%4.04%−2.10%0.12%

W2 / Q2: Independent contribution of executed-transition feedback. Table 4 retains the transition-based candidate representation, current state summary, recurrent architecture, and local candidate attention. Only the feedback signal changes: Transition reuses the selected transition embedding; Zero supplies a fixed zero vector; Action supplies an action embedding from pre-decision features. All three variants are independently trained with five seeds under the same protocol. Each variant uses its assigned feedback rule throughout training and evaluation. Table 4 reports greedy Gap (%; mean ± standard deviation (SD)) at each task's largest test size. Transition lowers mean Gap by 0.16–0.64 pp relative to Zero and 0.09–0.28 pp relative to Action across the six tasks. The effect persists with candidate inputs and recurrence retained. Sections 4.2–4.3 clarify that feedback supplies an explicit committed-change input alongside the compressed state summary, reusing only the selected candidate's pre-attention embedding.

Task / test sizeTransition ↓Zero ↓Action ↓
SMSP 5001.18 ± 0.101.60 ± 0.151.39 ± 0.12
UPMSP 100×67.48 ± 0.188.04 ± 0.257.74 ± 0.21
PFSP 100×5−0.16 ± 0.050.06 ± 0.07−0.04 ± 0.06
FFSP 100×122.27 ± 0.162.43 ± 0.202.36 ± 0.18
JSSP 100×202.22 ± 0.122.86 ± 0.172.50 ± 0.15
FJSP 40×100.39 ± 0.050.68 ± 0.080.53 ± 0.06

We sincerely appreciate the time and care you have devoted to our paper. Your constructive suggestions helped us strengthen the baseline comparisons and isolate the contribution of executed-transition feedback. Thank you again for your thoughtful review and valuable guidance.

Thanks to the author for the response; I am maintaining my positive rating.

Thank you very much for taking the time to consider our response and for your continued positive assessment. We sincerely appreciate your thoughtful comments and constructive suggestions, which have helped us clarify the contribution and strengthen the experimental evidence.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response to Reviewer i8dx: Quality–Cost Evidence, Matched Budgets, and Reproducibility

Dear Reviewer,

Thank you for your careful review and suggestions on fairness and reproducibility. We added feedback-controlled representation comparisons and matched-budget evaluations, and provided the original model checkpoints. Changes within the re-uploaded revised PDF are highlighted in blue.

W1: Practical benefit beyond symbolic completeness. Section 4.1 distinguishes the symbolic reconstruction guarantee from the empirical quality–cost benefits of edit encoding. The Table 5 analysis clarifies that Primary remains symbolically complete under deterministic reconstruction. Transition-complete (TC) encoding's 0.19–0.23 percentage-point advantage over Primary supports the value of explicitly encoding derived events. To assess quality–cost behavior independently of feedback, Table S13 (Appendix E.3) reuses zero-feedback TC and full-successor (FS) checkpoints with the same recurrent/local-attention backbone on identical flexible job-shop scheduling problem (FJSP) instances. TC achieves 0.12–0.14 percentage points (pp) lower Gap alongside 1.23–1.30× speedup and 17%–25% lower peak GPU memory allocation. Latency includes feature preparation, compilation or successor construction, inference, selection, and schedule execution. These results support a joint quality–cost benefit with feedback controlled.

Metric10×520×1040×10
FSTCFSTCFSTC
Obj. ↓106.28106.14197.06196.82368.92368.45
Gap (%) ↓10.3410.200.550.430.810.68
Latency (ms) ↓5404402350182048003700
Peak GPU memory (MiB) ↓5344786010478
Speedup (×) ↑1.23×1.29×1.30×
Memory reduction (%) ↑17.023.125.0

W2 / Q1: Matched comparisons with ReSched and MP-ASIL. We evaluate our Transition-Aware Dual-Scale Solver (TADS) against the requested ReSched and MP-ASIL, testing both ReSched variants (REINFORCE/PPO) and adding DANIEL, on identical FJSP instances and hardware under matched wall-clock inference budgets. Each method follows its specified decoding strategy without additional local search; only complete feasible schedules finished within the limit are eligible (Appendix D.6). Table S19 (Appendix E.7) controls training resources at 10/20 GPU-hours per run. All methods train at 10×5 with the same distribution and validation instances, retain their architectures and optimization procedures, and select checkpoints by validation performance within the budget. Evaluation uses 100 seconds per instance, without test-specific fine-tuning. TADS has the lowest mean Gap in all six settings, improving over MP-ASIL by 0.20–0.58 pp. Gap (%) is reference-relative; lower is better. Figure 6 uses the 20-GPU-hour checkpoints; TADS has lower Gap than the four baselines over the displayed time ranges. Table S17 (Appendix E.6) compares public FJSP instances under 100-second budgets. Checkpoints are selected without public data. TADS has the lowest mean Gap in these four groups. Mean Gap reductions over MP-ASIL are 1.18, 0.47, 0.42, and 0.06 pp across these groups. Together, these results support TADS's solution-quality advantage under matched computational budgets and its generalization across FJSP sizes and public benchmarks.

Method10 GPU-hours20 GPU-hours
10×520×1040×1010×520×1040×10
DANIEL6.90%−0.55%1.35%5.65%−1.10%1.07%
ReSched-REINFORCE6.40%−0.85%0.95%5.18%−1.38%0.69%
ReSched-PPO6.18%−1.02%0.78%4.92%−1.55%0.51%
MP-ASIL6.05%−1.20%0.60%4.62%−1.78%0.32%
TADS (Ours)5.65%−1.52%0.37%4.04%−2.10%0.12%
MethodBrandimarteHurink edataHurink rdataHurink vdata
DANIEL9.25%8.90%4.85%0.65%
ReSched-REINFORCE6.55%8.02%4.98%0.77%
ReSched-PPO6.30%7.73%4.77%0.69%
MP-ASIL7.10%7.05%4.54%0.48%
TADS (Ours)5.92%6.58%4.12%0.42%

W2 / Q2: Access to the original model checkpoints. The original model checkpoints are available through an anonymous link. These weights enable inference without retraining. Appendices D.2 and D.6 describe checkpoint selection and the evaluation protocol, respectively. Anonymous checkpoint link: anonymous.4open.science/r/industrial-scheduling-checkpoints-4CEC

We sincerely appreciate the time and care you have devoted to assessing our experiments and reproducibility materials. Your comments guided the matched comparisons and the addition of trained-model checkpoints. If these additions address your concerns, we would be grateful if you would consider revisiting your overall assessment in light of the new evidence.

AuthorsOpenReview ↗

General Response to the Area Chair and Reviewers

We sincerely thank the three reviewers for their careful assessment and constructive suggestions, and the Area Chair for their time and consideration. The reviewers' feedback has helped us clarify the contribution and strengthen the evidence for our Transition-Aware Dual-Scale Solver (TADS). All revisions are highlighted in blue in the re-uploaded revised PDF. We have added five experimental tables and Figure 6, simplified the Methods, clarified the roles of completeness and transition feedback, and expanded the descriptions of experimental protocols and results, as detailed in our individual responses. In response to the questions about transition representations, Section 4.1 now clearly separates exact symbolic reconstruction from the empirical value of learned change encoding. The reconstruction result provides a verifiable interface for sparse candidate descriptions that share unchanged state. Table 5 compares transition-complete (TC), Primary (all primary-field writes), Timing (timing edits only), and full-successor (FS) inputs under zero feedback. TC and Primary are both symbolically complete; their comparison isolates the value of explicitly exposing derived events, while Primary versus Timing tests broader edit coverage. Reusing the zero-feedback TC and FS checkpoints (TC-Zero/FS-Zero), Table S13 (Appendix E.3) shows 0.12–0.14 percentage points (pp) lower mean Gap, 1.23–1.30× speedup, and 17%–25% lower peak allocated GPU memory for TC. In the tables, SD denotes standard deviation, ms milliseconds, MiB mebibytes, and Obj. objective value. Independently, Table 4 retains the current state summary, TC candidate representation, recurrent architecture, and local attention, changing only the feedback signal across three variants trained with five seeds. Transition feedback lowers mean Gap by 0.16–0.64 pp versus zero feedback and 0.09–0.28 pp versus conventional action feedback at each task's largest test size across six formulations: single-machine scheduling problem (SMSP), unrelated parallel-machine scheduling problem (UPMSP), permutation flow-shop scheduling problem (PFSP), flexible flow-shop scheduling problem (FFSP), job-shop scheduling problem (JSSP), and flexible job-shop scheduling problem (FJSP). These controls distinguish the value of candidate change encoding from the additional benefit of feeding back the committed transition. To strengthen the fairness of external comparisons, we evaluate the requested ReSched and MP-ASIL, include both the REINFORCE and proximal policy optimization (PPO) variants of ReSched, and additionally compare DANIEL. The new FJSP comparisons use identical instances and hardware under matched wall-clock inference budgets (Appendix D.6). Figure 6 reports solution quality over elapsed time. Table S19 (Appendix E.7) additionally matches training resources at 10 and 20 GPU-hours per run, with checkpoints selected by validation within each budget and evaluation under a fixed 100-second inference budget. TADS has the lowest mean Gap in all six training-budget/test-size settings, improving over MP-ASIL by 0.20–0.58 pp. Table S17 (Appendix E.6) extends the 100-second comparison to Brandimarte and the three Hurink groups, where TADS also obtains the lowest mean Gap in each group. To improve readability, the Methods now foreground the compile–compare–execute–reuse workflow. Section 4.1 is reduced from five equations to two, with encoding and pooling details moved to Appendix B.3.1. We have clarified the timing, decoding, and checkpoint-selection protocols, and made the original model checkpoints available through an anonymous link, as described in our reproducibility response. Together, these revisions provide a clearer account of the design and direct evidence for its practical value under controlled conditions. We sincerely thank the reviewers again for their thoughtful feedback and valuable guidance, which have helped us clarify the contribution and strengthen the experimental evidence. We also thank the Area Chair for their time and consideration.

Dear Reviewer,

Thank you again for your careful review and constructive feedback.

As the rebuttal discussion period is approaching its deadline, we would like to kindly draw your attention to our detailed response and revised manuscript. We have directly addressed your main concerns through additional controlled experiments, fair comparisons with stronger neural scheduling baselines, and the provision of original model checkpoints for reproducibility. These revisions provide further evidence for the effectiveness and reliability of our approach.

We sincerely hope you will have an opportunity to review the new evidence before the discussion closes. If you find that our revisions have satisfactorily addressed your concerns, we would be very grateful if you could reconsider your overall rating.

We would also be happy to clarify any remaining questions.

Thank you very much for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.