acceptodds
Under review as a conference paper at ICLR 2027

Post-Calibration Topology Drift Reveals Operator-Dependent Numerical Sensitivity in Quantized Graph Neural Networks

Abstract

Post-training quantization (PTQ) is attractive for graph neural networks (GNNs), yet it is commonly evaluated on the same topology used for calibration. We study whether post-calibration topology drift changes the numerical discrepancy between a quantized GNN and its full-precision counterpart beyond the error already present on the calibration graph. We introduce Component-Preserving Degree-Preserving Rewiring (CP-DPR), a controlled intervention that changes neighborhood composition while preserving the node set, features, labels, model weights, edge count, exact node-wise degree sequence, and connected-component partition. We measure the baseline-adjusted excess logit divergence \(EQD_Z = QD_Z(A_\delta) - QD_Z(A_0)\). Experiments cover Cora, CiteSeer, PubMed, and Roman-empire; GCN and GraphSAGE; W8A8 and W4A4 simulated PTQ; five training seeds; and ten perturbation seeds per severity. At the strongest perturbation level, 10 of 12 external dataset–operator–precision settings exhibit positive excess divergence, with substantially larger absolute effects under W4A4. Roman-empire with GCN provides a consistent negative case that persists after validation-only model refinement. Activation diagnostics show negligible additional clipping in the dominant positive regimes and a stronger association with rounding mismatch. GraphSAGE further exposes an observer-blindness failure mode in which min–max calibration parameters remain unchanged despite internal activation-distribution motion. These results provide evidence that topology drift induces operator-dependent numerical sensitivity not captured by calibration-graph accuracy alone, while the evidence for operator dependence currently rests on one heterophilous benchmark.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph neural networks, post-training quantization, topology drift, low-precision inference, graph robustness, structural distribution shift

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_6hBcOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how changes in graph topology after PTQ calibration affect the numerical difference between quantized and not quantized GNNs. The authors propose CP-DPR, a rewiring method that preserves some graph properties, and EQDZ, a metric to measure the numerical divergence after topology changes. The experiments show that topology changes often increase divergence especially with 4-bit quantization. Roman-empire with GCN shows the opposite result.

Critical strengths
  1. The paper asks a precise and concise but at the same time interesting research question.
  2. The crossed bootstrap is appropriate for the experimental setup. The analyses of rounding, clipping, and recalibration are also useful, especially the observation that activation distributions can change without affecting min–max calibration parameters.
Critical weaknesses
  1. I found the paper difficult to follow, especially at the beginning. The authors start introducing formulas already in the abstract, without defining all the terms. In my opinion, the abstract and introduction should first explain the story of the paper. Introducing formulas with undefined symbols so early makes it harder to understand what the authors want to show. Moreover, for a reader who is not very familiar with quantization, some technical terms are used without enough explanation.
  2. My biggest concern is on EQDZ metric. QDZ normalizes the FP32–PTQ logit difference by the norm of FP32 logits. However, this norm can change when the graph topology changes. For example EQDZ could increase because the FP32 logit norm becomes smaller, even if the absolute quantization error stays the same.I suggest reporting: i) the absolute FP32–PTQ logit difference before and after perturbation; ii) how the FP32 logit norm changes with perturbation amount; iii) results with normalization fixed to the original graph.
  3. The paper shows changes in numerical divergence, but it is less clear if these changes affect actual predictions. I think there was enough space in the paper to investigate additional related research questions and make the work more complete and interesting.
  4. The paper reports larger EQDZ values for W4A4 than W8A8. However W4A4 already has a higher quantization error before perturbation. On Cora the relative increase compared to baseline is about 19% for W8A8 and 16% for W4A4; W4A4 has larger absolute divergence, but it is not clear if it is more sensitive in relative terms. Reporting baseline normalized EQDZ for all datasets and separating absolute and relative sensitivity can be useful.
  5. The opposite behavior on roman empire is interesting but the evidence comes from only one heterophilous dataset. It is difficult to know if this is caused by the architecture, heterophily, or something specific to roman empire. Testing more heterophilous datasets can be useful.
  6. In the introduction, the authors motivate the problem by explaining that real graphs change over time. I understand the choice of a controlled rewiring method, since it helps to isolate specific numerical effects. However, given this motivation, I think it would also make sense to test the method on temporal graph datasets, where the topology actually changes over time.
  7. CP-DPR preserves degrees and connected components, but the are others structural properties that can vary (for example homophily, and this is connect to W5). Analyzing which graph properties are most related to divergence could help explain the results.
Questions for authors

See weaknesses.

Reviewer_rq72OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies about how to retain robustness of quantized graphs for when input graph topology is perturbed, by identifying what causes the error between the full-precision GNN and the quantized counterpart. The paper proposes a metric EQDzEQD_z , and a controlled setup protocol, CP-DPR to measure the topology-induced error separated from quantization effects. From the actual use cases, the paper states that quantized GNNs should not be only evaluated with its calibrated graphs, and the quantizer modules need additional designs to cover topological drifts.

Critical strengths
  1. The paper explores an interesting research direction: topology-aware robustness in quantized GNNs.
Critical weaknesses
  1. The motivation and benefits for analyzing the targeted problem using the proposed method is unclear.
    • It seems that robustness to topology drifts seem to be a property that can be remedied with inductive learning (where the GNN is trained to inference unseen topology), but the paper does not clarify how these concepts differ.
    • There are no baselines to compare with the suggested method, so it is unclear whether the metric is working well.
    • Also, the use-case range for the analysis is too limited. More graph datasets with larger size and GNN architectures (e.g., GAT, GIN) is recommended for generalization of the trends.
  2. The presentation of the problem and proposed method requires substantial revision:
    • The experimental results are poorly organized, leaving readers to infer the intended conclusions from EQD_z, a metric that is not clearly explained.
    • Some conclusions rely on a narrow experimental setting(Roman-empire, GCN), raising concerns about their generalizability.
    • The paper provides insufficient background on existing approaches to errors caused by topology drift, making it difficult to assess the proposed method’s advantages.
Questions for authors
  • Can topology-induced error be remedied with inductive learning? if not, why?
  • Can EQD_z be compared with other baseline metrics used for full-precision GNNs? Providing strong evidence for why EQD_z is a good representation for measuring errors incurred by topology drifts in quantized inference is recommended.
  • Prior work appears to address similar problems in full-precision GNNs. Could the paper compare its metrics and evaluation protocol with those studies qualitatively and clarify which challenges remain unaddressed, or is different from existing approaches/problems?
  • Do the reported EQD_z trends and conclusions hold across larger graph datasets and other GNN architectures, such as GAT and GIN?
Reviewer_iWzbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks whether PTQ GNNs calibrated on a reference graph show extra FP32–PTQ discrepancy when the topology later changes. It proposes CP-DPR, a degree- and component-preserving rewiring, and a baseline-adjusted metric EQD_Z. Experiments use 4 small datasets, GCN/GraphSAGE and W8A8/W4A4. EQD_Z is mostly positive, but negative for Roman-empire+GCN.

Critical strengths

The question is clearly scoped, and EQD_Z is a sensible way to separate topology-induced error from baseline quantization error.

The crossed-bootstrap statistics are careful, and the limitations (including the negative case) are reported candidly.

Critical weaknesses

“Post-calibration” is not isolated. Weight quantization does not depend on G₀, yet its error propagates through Â, so it changes with topology anyway. Recalibration on G_δ also does not reliably help. A weight-only vs. activation-only ablation is needed to show that frozen calibration is actually the cause.

The evidence is incomplete. Baseline QD_Z(A₀) is missing for all external settings. The claim that component preservation makes little difference is not backed by any numbers. Mechanism correlations are pooled across severities, which inflates them because both quantities grow with δ_E.

The manuscript looks unfinished. The main text is about 6 pages with no figures. Appendix B contains garbled text, the reproducibility statement has leftover template text, there is no code, and several passages refer to “the submission version”.

Questions for authors

What is EQD_Z under weight-only and activation-only quantization?

How do homophily, FP32 accuracy and ‖Z_FP‖ change under rewiring, and does the Roman-empire sign hold for the unnormalized error?

What are the QD_Z(A₀) values for all settings, and the within-severity correlations?

Does the effect appear under realistic temporal or edge-addition drift?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.