Writing study: ProGMLP: A Progressive Framework for GNN-to-MLP Knowledge Distillation with Efficient Trade-offs
← AAAI 2026 oral writing corpus
Weigang Lu, Ziyu Guan, Wei Zhao, Yaming Yang, Yujie Sun, Zheng Liang, Yibing Zhan, Dapeng Tao · Official AAAI 2026 oral record
Figure, table and section ledger
Editorial contribution type: Training, objectives & optimization (primary).
Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.
- Canonical PDF
- 9 physical pages
- Before references / appendix
- ≈ 7.14 page-equivalents
- Main figures / tables
- 6 / 4
- Attached appendix figures / tables
- 0 / 0
- Attached appendix span
- No boundary detected; not assumed zero
Extraction notes: No attached appendix boundary detected. Zero means no recognized caption, not independently verified absence.
Source-bound correction: Checked 2026-10-25 against the manuscript scan and physical-page bounding boxes. Restore unnumbered main headings, including the separate Motivation before Methodology. Figures 1–6 and Tables 1–4 are distinct labels. The extractor finds the identical Figure 1 candidate twice at the same coordinate; normal label de-duplication already removes that duplication.
Section-by-section lengths and source pages
The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.
All detected figure and table locations
Main text
Attached appendix
No recognized caption labels.
Argument and organization
- Introduction; Preliminaries; Related Works; Motivation
Moves from a fixed accuracy-cost operating point to the need for interruptible inference, then distinguishes GNN-to-GNN, GNN-to-MLP and ensemble predecessors before defining the proposed intervention.
- Methodology; Train and Inference; Complexity
Explains a sequence of student MLPs, progressive distillation and mixup, followed by the confidence-based stopping rule and separate training/inference complexity statements.
- Experiments; Conclusion
Tests the accuracy-versus-executed-students curve before peak performance, large-graph runtime, inductive transfer, component ablation and parameter sensitivity.
Problem, claim and contribution
The opening frames a deployment choice rather than only higher node-classification accuracy: a fixed MLP cannot select different accuracy-cost operating points at runtime. The separate Motivation section distinguishes proactive early exit from interrupted anytime inference, giving the later stopping rule a concrete purpose.
Abstract and Introduction, physical pp.1-2; Motivation, p.3 · Main paper
Argument and section order
The method is ordered by dependency: student inputs and inherited parameters, the distillation objective, mixup augmentation, then inference and complexity. Experiments return first to the stated trade-off using executed students as the horizontal variable, before reporting peak accuracy and measured milliseconds separately.
Methodology, pp.3-5; Figure 3, p.5; Main Results, p.6; Table 3, p.7 · Main paper
Figures, captions and visual explanation
Figure 2 places a common pretrained teacher over two successive student stages and uses an explicit legend for loss weights, mixing ratios and hidden/prediction projections. Figure 5 uses aligned small-multiple ablation panels; Figure 6 assigns different markers as well as colors to parameter sweeps. The overview is explicitly a training architecture, not the inference stopping procedure.
Figure 2, physical p.3; Figures 5-6, p.7; rendered pages inspected · Main paper
Tables, comparisons and evidence
Table 3 puts accuracy and inference time beside each other for two large graphs and preserves GAT's out-of-memory result instead of treating it as zero. Table 4 separates inductive evaluation from the main comparison. The executed-student count in Figure 3 is a cost proxy; it should not be relabeled as elapsed time or energy.
Figure 3 and hardware/setup, p.5; Tables 3-4, p.7 · Main paper
Appendix and reproducibility
The published PDF carries the architecture, objectives, stopping rule, hardware and hyperparameter search ranges in the main text and provides a code URL on its first page. It has no attached appendix in the inspected file. A future drafting plan should retain these defining details in the main paper and link exact implementation protocols rather than assume a code link explains them.
Code link, p.1; Methodology and Experimental Setup, pp.3-6; acknowledgments/references, pp.8-9 · Main paper
Clarity and writing conventions
The repeated PTS, PKD and PMA names connect motivation, method subsections and ablation labels. Metric wording needs equal care: the improvement rows report relative changes, whereas the underlying accuracies are percentages; for example 80.06 to 87.95 on Pubmed is 7.89 percentage points, not a 9.85-point increase.
Methodology, pp.3-4; Table 4 and Inductive Setting, p.7 · Main paper
What to reuse—and what not to infer
Transferable lesson: Turn a flexible-compute claim into a matched accuracy-cost curve, then report actual runtime separately. Use one training diagram to expose teacher/student signals and a separate equation or procedure for inference stopping.
Boundary: Executed-student counts are not wall-clock measurements, and the displayed accuracy dispersion is not automatically a confidence interval. Preserve out-of-memory entries and relative-versus-absolute improvement units. This writing study does not establish deployment performance on other hardware or validate the implementation.
Apply this lesson using the argument and evidence worksheet.
Inspection record and sources
- Checked
- 2026-10-25
- Reading status
- Writing-focused manuscript scan + selected visuals inspected
- Main text
- Writing-focused text scan of physical PDF pp.1-7, including the framing, related work, staged training and inference definitions, experiments and conclusion. Physical pp.3 and 7 were rendered and visually inspected; the acknowledgments/reference boundary on p.8 was checked. This is not proof verification, implementation validation or experimental reproduction.
- Appendix / supplement
- The inspected nine-page proceedings PDF ends with acknowledgments and references on pp.8-9; no attached appendix or main-text appendix pointer was identified. The linked code repository and any other supporting material were not inspected, so their existence or contents are not inferred from this PDF.
- Visually inspected PDF pages
- 3, 7
This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.