acceptodds
Withdrawn from ICLR 2027

SchedGPT: Towards The First Auto-Regressive Foundation Model for List Scheduling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

List scheduling underlies many combinatorial problems in manufacturing and distributed computing, yet existing learning-based schedulers are typically specialized to one problem family and one architecture. We introduce SchedGPT, an autoregressive foundation model with a graph encoder that handles diverse list-scheduling variants within a single architecture. The key insight is a unified heterogeneous graph representation over task and processor nodes, connected by precedence and eligibility relations, plus a unified priority-list encoding that specifies task ordering and per-task processor assignment and is deterministically decoded into a schedule. SchedGPT combines a relation-aware GATv2 encoder with a nanoGPT autoregressive decoder, factorizing each construction step into task sequencing and nested processor assignment via two pointer heads. Trained from scratch with a rollout-baseline policy-gradient objective, SchedGPT outperforms specialized state-of-the-art algorithms on manufacturing scheduling (JSSP and FJSP) and workflow scheduling (five variants), demonstrating that list-scheduling problems can be unified under one architecture.

Primary area:
optimization
Keywords:
Fundation Model, Optimization, List Scheduling

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_SLmMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SchedGPT, a unified autoregressive learning framework for several list-scheduling problems, including Job Shop Scheduling (JSP), Flexible Job Shop Scheduling (FJSP), and Workflow Scheduling (WSP). The key idea is to represent different scheduling instances using a common heterogeneous graph consisting of task and processor nodes, and to encode scheduling solutions as aligned task-priority and processor-assignment lists. A deterministic schedule builder then converts these lists into feasible schedules. The model is trained using policy gradients with a rollout baseline to minimize makespan. Experiments on multiple scheduling benchmarks demonstrate competitive performance against specialized learning-based schedulers, particularly when the model is trained separately for each problem family.

Critical strengths
  1. Meaningful problem and unified formulation. The paper proposes a common learning framework for scheduling problems that have traditionally been studied separately. The unified task–processor representation and priority-list-based solution construction provide a coherent approach to accommodating different precedence structures and processor compatibility constraints.

  2. Reasonable and well-structured framework. The proposed architecture integrates graph representations, autoregressive decision-making, and deterministic schedule construction in a natural manner. In particular, separating learned task–processor decisions from feasibility-preserving schedule construction is a sensible design choice. The overall methodology is clearly motivated and relatively easy to follow.

Critical weaknesses
  1. Insufficient evidence for the foundation-model and generalization claims. The strongest results are obtained by SchedGPT-S, which uses separately trained parameters for different problem families. In contrast, the jointly trained SchedGPT-U exhibits noticeable performance degradation and falls behind the strongest specialized baselines across the evaluated domains. Additional experiments on unseen scheduling variants, cross-domain adaptation, or fine-tuning efficiency would make the foundation-model positioning more convincing.

  2. Incomplete validation of some important design choices. The current evaluation would benefit from more targeted ablations examining the contributions of the autoregressive GPT decoder, heterogeneous graph encoder, and engineered dynamic features.

  3. Limited methodological novelty relative to existing work. While the overall integration is meaningful, several key components, including heterogeneous task–processor graphs, priority-list-based schedule construction, attention-based task and machine selection, and autoregressive neural combinatorial optimization, have been explored in prior studies. The paper's main novelty therefore appears to lie primarily in combining these ideas into a common framework spanning multiple scheduling families. A clearer discussion of the technical advances beyond these existing approaches would strengthen the contribution.

Questions for authors
  1. Could the authors further clarify the main methodological differences between SchedGPT and existing approaches such as HGNN, ReSched, and priority-based neural DAG scheduling? In particular, which aspects of the unified representation and autoregressive construction are fundamentally different from existing scheduling methods? It would also be helpful to clarify the conceptual advances over unified neural combinatorial optimization frameworks such as RouteFinder.

  2. The results suggest that SchedGPT-U can accommodate multiple scheduling families, but with a performance reduction relative to SchedGPT-S. Have the authors investigated whether joint training provides benefits beyond architectural unification, such as improved sample efficiency, transfer to unseen problem variants, or faster adaptation through fine-tuning? Such results would help substantiate the foundation-model claim.

  3. How much improvement is attributable specifically to the autoregressive GPT decoder compared with a simpler state-conditioned pointer network or non-autoregressive alternative? It would also be useful to report inference latency and computational scaling with the number of tasks, especially relative to one-shot priority-based scheduling methods.

  4. Dynamic features for general DAG workflows. Appendix B.2 defines the frontier-operation set and several progress-related features using job-specific operation indices, which appear naturally suited to JSP/FJSP. However, general workflow DAGs may contain multiple simultaneously ready successors and do not necessarily admit the same job-chain representation. Could the authors clarify how these dynamic features are defined and implemented for WSP? Are any domain-specific feature adaptations required?

I would be open to reconsidering the rating if the authors can provide convincing evidence of cross-domain transfer or adaptation benefits and clarify the implementation and methodological concerns above.

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

This paper proposes SchedGPT, which is claimed to be the first autoregressive foundation model for list scheduling. SchedGPT can handle with multiple list scheduling problems without changing the model and network architecture. As the representation foundation of SchedGPT, a uniform mathematical formulation for list scheduling is formalized, and a unified heterogeneous graph presentation for tasks and processors is described. As the model architecture, SchedGPT introduces a GATv2 encoder and a nanoGPT autoregressive decoder for task sequencing and processor assignment. Experiments are conducted on on JSP, FJSP and WSP, and the results show that SchedGPT-S, which is the problem-specific-trained version of SchedGPT, outperforms many specialized SOTA methods.

Critical strengths
  1. This paper seeks to solve the critical problem that existing learning-based schedulers are typically specialized to one problem family and one architecture. SchedGPT can handle with multiple list scheduling problems without changing the model and network architecture.

  2. This paper provides a uniform mathematical formulation for list scheduling. This provides the mathematical foundation for proposing a model that can adapt to and generalize various scheduling problems.

  3. Experiment results on various scheduling problems show that SchedGPT-S outperforms many specialized SOTA methods.

Critical weaknesses
  1. Although this paper describes the components and computation procedures of SchedGPT in much detail, many designs in this paper, such as the heterogenous graph representation, GATv2 encoder and nanoGPT decoder, are existing methods, and the paper doesn’t adequately explain why such designs can achieve SchedGPT’s adaptability in various list scheduling scenarios.

  2. There is no ablation experiments about how each component in SchedGPT achieves SchedGPT’s optimization performance and adaptability in various list scheduling scenarios.

  3. This paper claims that SchedGPT can handle diverse list-scheduling variants within a single architecture. However, there are many other studies that can achieve this. For example, both Jeon et al., 2023 and Qi et al., 2025 have conducted experiments on both WSP and JSP. So, the contribution of this article is not significant enough. Meanwhile, the list scheduling is not limited to be applicable to JSP, FJSP or WSP that are discussed in this paper. There is also parallel machine scheduling problem (PMSP), resource-constrained project scheduling problem (RCPSP), and their variations, among others. Therefore, the author should thoroughly discuss the boundaries of the applicability of the proposed SchedGPT.

  4. The experimental results show that the SchedGPT-U cannot outperform most learning-based methods. Only the domain-specific-trained SchedGPT-S can generate competitive results compared with most learning-based SOTA methods. This indicates that the generalizability of SchedGPT still needs to be improved.

  5. Autoregressive scheduling models are known to be time-consuming, but this paper doesn’t conduct complexity analysis or report runtime of SchedGPT.

Questions for authors
  1. Which design in SchedGPT has made the greatest contribution to its adaptability in various list scheduling scenarios? The unified instance and solution representation, the GATv2 encoder, the nanoGPT decoder, or the training method?

  2. Is SchedGPT applicable to the other combinatorial optimization problems that involve generating a priority list, such as ranking and routing problems?

Reviewer_i7YiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

To address the fragmentation of learning-based schedulers across different scheduling problems, this paper proposes a unified autoregressive framework called SchedGPT. The method represents scheduling instances as heterogeneous task–processor graphs and constructs aligned task-ordering and processor-assignment lists using a graph encoder and a Transformer decoder with two pointer heads. A deterministic schedule builder ensures feasibility, while policy-gradient training optimizes makespan without optimal-solution labels. Experimental results on job-shop, flexible job-shop, and workflow scheduling demonstrate competitive performance under domain-specific training and the feasibility of using a single jointly trained model across these problem families.

Critical strengths

1.The paper is clearly motivated: representing manufacturing and workflow scheduling within a unified task–processor graph framework helps reduce fragmentation in model interfaces across different scheduling problems.

2.The method’s architecture is well aligned with the scheduling constraints.

3.SchedGPT-S outperforms the listed baselines on the aggregate metrics for JSP, FJSP, and Pegasus, suggesting that the unified architecture is applicable across multiple scheduling settings rather than being restricted to a narrow scenario.

Critical weaknesses

1.Ablation studies on the core components are missing. The experimental results are primarily presented in Tables 1 and 2, and the appendix provides no component-level ablations, making it difficult to identify the sources of the performance gains.

2.The main performance advantages come from the separately trained SchedGPT-S rather than SchedGPT-U, which shares parameters across domains. SchedGPT-U underperforms the strongest baseline in each of the three domains and performs substantially worse than SchedGPT-S. Therefore, the claims regarding its foundation-model capabilities and cross-domain generalization are not yet sufficiently supported.

3.Page 2 states that the trajectory log-probability and entropy are normalized by the number of tasks. However, Equation (23), using the summed quantities defined in Equations (19) and (20), does not explicitly include this normalization, and Algorithm 2 does not clarify it either. The authors should reconcile the description with the implemented loss.

4.An efficiency analysis is missing. Page 10 states that the appendix reports computational costs, but no corresponding measurements are provided in the current appendix.

Questions for authors

1.Please clarify whether SchedGPT-S and SchedGPT-U use comparable per-domain training data volumes and computational budgets, and analyze the performance degradation of SchedGPT-U. Is there evidence of benefits in cross-domain transfer or data efficiency?

2.Does the implemented loss include 1/n1/n trajectory-length normalization? Please clarify.

3.Please provide key ablation studies, particularly comparisons with a variant without the history Transformer and a simpler policy using the same dynamic features and schedule builder, as well as an evaluation of the normalization strategies.

4.What is the actual inference cost? Please provide an efficiency analysis, including end-to-end runtime covering graph encoding, dynamic feature computation, decoding, and schedule construction, and compare it with the main baselines.

5.Please provide additional experimental configurations and statistical information, including training instance sizes, domain mixing ratios, the sources of baseline results, results across multiple training seeds, and the complete sampling budget and baseline comparisons supporting the reported 3.71% gap.

Reviewer_mLm7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work addresses the list scheduling problem in the manufacturing industry and computational workflows, using a unified framework for various tasks. It proposes a framework that uses heterogeneous graph encoder with an autoregressive decoder to generate task ordering and processor allocation, and a deterministic builder to compile schedules. The policy is trained with REINFORCE using makespan as feedback. The model has two variants: 1/ trained for problem-specific models separately 2/ trained as a single model jointly across all problems. The paper compares model performance with standard baselines on JSP, FJSP, and 5 Pegasus WPS datasets. The separate trained model has the lowest reported average makespan gaps, while jointly trained models fall behind several baselines.

Critical strengths

The paper addresses an important resource-allocation problem that balances resource utilization and overall computational efficiency.

It follows a coherent progression from problem definition to solution, architecture to training, with overview figures that help readers to follow the formulation.

The evaluation includes multiple public benchmarks and established baselines. It compares different training settings (task-separate v.s. joint) that provides insights on unified architecture and parameter sharings.

Critical weaknesses

Incomplete related work. Relevant approaches using pretrained LLM are not included. The paper can include the comparison and justify the choice of training a specialized architecture from scratch.

The paper claims to be the first autoregressive foundation model with a graph encoder for list scheduling. However, it misses component ablations and cost analysis. The benefits of GATv2 encoder and explicit Transformer are not isolated through comparisons with simple encoders or current-state policy.

The paper has an incremental contribution over ReSched as ReSched already supports multiple scheduling variants through common architecture. This work additionally supports WSP, and a two-stage decision requires clear justification.

The paper claims to be a foundational model, however it shows limited evidence for the benefit of joint training. SchedGPU-U underperforms than specialized baselines, while comparisons controlling model capacities and task families are not established, limiting the conclusion about cross-task transfer and generalization.

Questions for authors
  1. Can you discuss how ScheGPT relates to a. LLMs can Schedule(2024):https://arxiv.org/html/2408.06993 b. ReflecSched(2025):https://arxiv.org/html/2508.01724v2 which explores pretrain LLM-based scheduling? It would be good to justify what motivates the work training from scratch rather than initialization or distillation from the teacher followed by RL.

  2. Could you explain the motivation of introducing graph encoder? What scheduling-information does graph structure help encode, and why GATv2 was selected. Have you compared it with naive Transformer encoder or relation-aware attention?

  3. The paper uses a history-model structure. Could you provide an ablation isolating the contribution of history relative to a policy conditioned on the current state? It would be helpful to show what is not captured by the dynamic features and how history representation closes the gap.

  4. In table 1 the improvements are sensitive to various groups, especially on CyberShake(<500). For WSP results, can you the 95% CI and breakdowns by workflow and instance size? Group-level medians or quantiles, results across multiple training seeds can better quantify results.

  5. The introduction states the log-p and entropy are normalized by number of tasks, whereas Eq. 19, Eq. 20 defines sum and Eq. 23 does not include the 1/n factor. Please specify where normalization occurs.

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

Related papers

Discussion (0)

Sign in to comment.