acceptodds
Under review as a conference paper at ICLR 2027

When Is Reweighting Not Enough? A Hodge Certificate for Rewiring Noisy and Incomplete Graphs

Abstract

Graph repair methods change an edge in one of two ways. They change how much it counts, which we call reweighting, or whether it exists, which we call rewiring. Structure learners mix the two moves and rarely ask which one a graph needs. We give an exact answer to one form of this question. The predictions of a frozen node classifier, compared across each edge, form a flow on the graph. Hodge decomposition splits the flow into a gradient part, which one score per node explains, and a cycle part, which no such score explains. We prove that no positive reweighting of a fixed edge set lets one score per node explain the cycle part, and that weights of spread γ move its share of the flow by at most a factor γ. A nonzero cycle part therefore certifies that an exact fit requires rewiring. We turn the certificate into the Hodge Router. It rewires an instance only when the cycle fraction is large relative to its chance level, ranks the edits by their share of the cycle part, and sends the edits on which the structural signals disagree to a language model under a fixed query budget. The bound holds on every weighting we draw. Corruption raises the diagnostic to about 1.6 times its clean value, and the diagnostic predicts which operation helps better than seven competing statistics, with a partial Spearman correlation of 0.38. Under random corruption the router matches or exceeds the better fixed operation on every dataset. Under Metattack it rewires nearly every instance, and under PRBCD it reweights the graphs on which reweighting wins, so it follows the operation each attack favors. It has the best average accuracy of all compared methods under both attacks, ahead of the joint structure learners JDR and Pro-GNN.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph neural networks, graph structure learning, Hodge decomposition, robust graph learning

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_9uHnOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies when graph repair should use reweighting versus rewiring. The authors construct an edge flow from frozen classifier predictions and use Hodge decomposition to separate gradient and cycle components. They prove that positive reweighting on a fixed support cannot eliminate a nonzero cycle component, and propose the Hodge Router, which uses the normalized cycle residual to choose between reweighting and rewiring. The rewiring branch further combines link scores, cycle contributions, and a bounded LLM-based arbiter. Experiments across random corruption, Metattack, and PRBCD show that the proposed diagnostic predicts which operation is preferable and that the router achieves strong robustness results.

Critical strengths

1.The paper asks a clear and relatively underexplored question: not how to repair a graph, but when reweighting is fundamentally insufficient and structural editing is preferable.

2.The theoretical formulation is elegant. The Hodge decomposition gives a clean interpretation of prediction inconsistency, and the reweighting result provides a principled connection between cycle residuals and support changes.

3.The experimental evaluation is extensive, covering multiple datasets, corruption mechanisms, adversarial attacks, competing diagnostics, graph-repair baselines, backbone transfer, sensitivity analysis, and computational cost.

4.The controlled comparison between reweighting and rewiring is a particular strength. Both operations use the same candidate support, backbone, and edit budget, making the routing analysis more interpretable.

Critical weaknesses

1.The theoretical certificate concerns fitting a fixed prediction flow, whereas the practical goal is downstream classification after retraining on the repaired graph. Since predictions can change substantially after repair, the connection between the theorem and the final learning objective remains indirect.

2.The impossibility result for exact fitting is somewhat narrow. Whether yy belongs to the gradient subspace is support-dependent regardless of weighting, so the more practically relevant contribution is the bounded-residual argument and its empirical predictive value rather than the exact-fit statement itself.

3.The diagnostic depends on the choice of the antisymmetric comparison matrix KK. The default class-frequency ordering is somewhat heuristic, and although alternative choices are evaluated, a stronger justification for why this construction should reflect graph corruption would improve the method.

4.The repair experiments assume an edit budget tied to the known perturbation rate. This is convenient for controlled comparison but may overestimate practical applicability when the amount of graph corruption is unknown.

5.The predictive relationship between the diagnostic and the preferred operation is meaningful but moderate rather than strong. A partial Spearman correlation of 0.38 suggests that substantial instance-level variation remains unexplained.

Questions for authors

1.Can the authors clarify the theoretical connection between fitting the frozen prediction flow and improving the classifier after retraining, where the prediction flow itself may change?

2.How sensitive is the routing decision to the choice of comparison matrix KK, and what principle motivates class-frequency ordering beyond its ability to generate non-gradient flows?

3.How would the Hodge Router determine the edit budget in a realistic setting where the perturbation or corruption rate is unknown?

4.The partial correlation between κdiag\kappa_{\mathrm{diag}} and the operation gap is 0.38. What are the dominant failure cases where the diagnostic routes an instance to the wrong operation?

5.To what extent does the performance gain of the full Hodge Router come from operation routing versus improved cycle-aware rewiring and semantic arbitration?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9uHn

Thank you for the thoughtful review. References below are to the proposed revised manuscript.

W1 / Q1: The link to a retrained classifier. The certificate concerns a specified flow on a fixed support; retraining may change that flow. We therefore use it to motivate a diagnostic, not to equate potential-fitting error with classification loss. The classification link is tested directly: kappa_diag has partial Spearman correlation 0.38 with the Rewire–Reweight gap, with 95% CI [0.21,0.52] (Table 1). The resulting gate matches or exceeds the better fixed policy on every dataset and has average oracle regret of 0.09 Macro-F1 points (Table 2). Table 5 verifies the fixed-flow bound using random weightings; it does not measure changes to the flow after retraining.

W2: Why emphasize the two-sided bound? The exact-fit condition is indeed membership in the gradient subspace fixed by the support. The quantitative contribution is rho_0/gamma <= R_W/||y||_W^2 <= gamma*rho_0: weights with bounded spread cannot reduce fixed-flow inconsistency arbitrarily. The bound becomes weaker as the permitted spread grows. The revision therefore foregrounds this quantitative statement and presents exact-fit impossibility as its corollary, with its scope separated from the empirical routing result.

W3 / Q2: Motivation and sensitivity of K. Proposition 1 requires K not to be a difference of class scores. The sign comparison is a simple order-based construction with that property; frequency ordering provides a training-data-based convention rather than relying on arbitrary label indices. Changes in frequencies that preserve the order leave the construction unchanged.

For one-hot predictions, nonzero circulation requires cycles involving at least three predicted classes. Figure 2 illustrates how a cross-class shortcut creates such a residual. This is a mechanism, not an assertion that every multi-class cycle is wrong. The evidence for relevance comes from the predictive and routing experiments.

For C=3, admissible K choices yield proportional cycle components up to sign on fixed predictions and support, preserving squared contribution rankings. This does not imply invariance of the normalized diagnostic or every routing decision. Table 16 directly tests alternatives: average accuracy changes are -0.15/-0.26/-0.93 points for prototype-cosine/native-order/random antisymmetric K under Metattack 15%. The evaluated alternatives thus remain within one point of the default.

W4 / Q3: Unknown corruption and the edit budget. The gate never reads the corruption rate. A budget is required only after an instance is routed to Rewire. Table 21 tests substantial misspecification under Metattack 15%: doubling the budget costs 0.54/0.53/0.37 points on Cora/CiteSeer/PubMed; halving it costs 1.82/1.47/1.45. At double budget the router still exceeds the strongest competing baseline on each graph. This establishes budget tolerance in that regime, rather than a general estimator of an unknown corruption rate.

W5 / Q4: Moderate correlation and weak regimes. The diagnostic is not a perfect instance-level predictor; the decision-level result is that it recovers most of the available routing benefit. Its average oracle regret is 0.09 points, versus 0.16 for uncalibrated rho and 0.30–0.49 for the other diagnostics (Table 1). Both kappa_diag and rho have partial-correlation intervals excluding zero; the other deployable diagnostics do not.

Table 7 locates the weakest association in betweenness deletion on Cora/CiteSeer (0.29/0.26, nonsignificant after Holm correction). At 15% deletion on Cora, Reweight scores 82.24 versus 79.74 for Rewire. Under Metattack, the correlations rise to 0.55/0.60/0.52 and are all significant. These strata characterize where the diagnostic is more or less informative without claiming an unmeasured breakdown of all individual misrouting errors. Under PRBCD, the gate correctly favors Reweight on 58/60 Cora/CiteSeer instances.

Q5: Which component contributes what?

  • Random corruption (Table 2): routing adds 0.04–0.20 points over the better fixed policy across the four datasets in Table 2. Arbitration adds another 0.31–0.37 on the three text graphs.
  • Metattack 15% (Table 16): removing cycle-aware ranking costs 5.57 points on average; removing the arbiter costs 1.36. With no arbiter, the residual gate is within 0.2 points of Always-Rewire on these three datasets, so the main gain here is how the graph is rewired.
  • PRBCD (Table 11): the full router averages 69.7, versus 68.4 for Always-Reweight and 67.6 for Always-Rewire. The preferred operation varies across datasets and rates, making adaptive selection useful. The full-router comparison also includes arbitration, so we do not attribute its entire advantage to routing alone.

These conditional comparisons distinguish operation selection from improvements within rewiring; they are not additive.

The rebuttal addresses most of my concerns and provides useful clarification on the scope of the theoretical certificate, the role and sensitivity of KK, edit-budget robustness, and the relative contributions of routing, cycle-aware rewiring, and arbitration. In particular, the additional discussion of oracle regret and the weak/strong diagnostic regimes improves my confidence in the empirical relevance of the proposed certificate. Some limitations remain, especially that the theoretical guarantee applies to a fixed prediction flow rather than directly to the retrained classifier. Overall, the response strengthens the paper, and I maintain my original rating.

AuthorsOpenReview ↗

Follow-up response to Reviewer 9uHn

Thank you for the supportive assessment. We are glad the clarifications were useful. Below we address the remaining point.

Some limitations remain, especially that the theoretical guarantee applies to a fixed prediction flow rather than directly to the retrained classifier.

Response. The guarantee is stated on a fixed flow because the router chooses the operation before either repair is trained. Its structural part also applies to the retrained flow (E1). The link to classification accuracy is measured empirically, with results supporting the decision across the evaluated backbones and flow constructions (E2–E4).

E1. What carries over after retraining. The flows that node potentials can fit form im⁡(B⊤)\operatorname{im}(B^\top), which is fixed by the candidate edge set EcE_c, not by the classifier or weights. Reweight keeps EcE_c, so retraining does not enlarge this subspace. Theorem 1 applies to every flow on EcE_c, including the retrained flow: weights of spread γ\gamma keep its residual share between ρ0/γ\rho_0/\gamma and γρ0\gamma\rho_0, where ρ0\rho_0 is that retrained flow's own unweighted share.

The theorem does not determine this new ρ0\rho_0 or classification accuracy. The frozen flow supplies a diagnostic available before repair; Sec. 4.2 states its link to classification as the hypothesis tested experimentally.

E2. Prediction of outcomes after retraining. The Sec. 5.3 operation gap is the Macro-F1 of the classifier retrained after Rewire minus that after Reweight. Across 370 instances, the partial Spearman correlation of κdiag\kappa_{\rm diag} with this gap is 0.38, with CI [0.21, 0.52]; no non-Hodge deployable diagnostic in Table 1 reaches 0.10. The gate is within 0.09 points of the per-instance oracle, which sees both retrained classifiers.

E3. Transfer across retrained backbones (Table 19). The flow, gate, and edits are computed once from the frozen GCN. Four backbones are trained from scratch on the resulting graphs under Metattack at 15%, pooled over Cora, CiteSeer, and PubMed. The policy ordering is the same for all four. GAT and GraphSAGE ignore edge weights, so their Always-Reweight results use the unweighted candidate support.

BackboneObserved graphAlways-ReweightAlways-RewireHodge Router
GCN70.8073.0678.8180.00
GAT70.8771.6278.0579.13
APPNP71.9373.8479.3780.46
GraphSAGE69.6471.0277.2478.31

E4. A flow without a trained classifier (Table 20). Replacing the frozen classifier with soft assignments to feature centroids removes the trained classifier from flow construction. Under random corruption, the residual gate matches or exceeds the better fixed policy on all four datasets and exceeds the random gate that rewires the same share of instances.

Macro-F1CoraCiteSeerPubMedDBLP
Always-Reweight80.2660.9583.2672.81
Always-Rewire79.8260.4283.8171.31
Random gate80.0560.7383.6072.55
Residual gate80.4661.1583.8572.90

Summary. The theorem characterizes what positive weighting can fit on a fixed candidate support, for both frozen and retrained flows. Whether the diagnostic predicts classification performance after retraining is an empirical question, tested in E2–E4 rather than guaranteed by the theorem.

Revision. The updated PDF is now uploaded. Blue text after Corollary 1 in Sec. 4.2 (page 5) clarifies that retraining on EcE_c does not enlarge the gradient subspace and that the theorem applies to the retrained flow. “Reading the certificate” in App. B.1 (page 18) explains why the gate reads the frozen flow, what carries over after retraining, and how Sec. 5.3 and Tables 19–20 evaluate downstream accuracy.

Thank you again for the constructive discussion and your support.

Reviewer_9pnmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on two questions in graph repair: Can we tell, before any repair is trained, whether reweighting could ever fit the evidence on the edges? And does the answer predict which operation improves accuracy? To answer the question, this paper formalizes graph repair as a choice between positive reweighting of a fixed support and editing of the support, and provides the formal proof. Hodge router is proposed to select between reweighting and rewiring based on normalized cycle residuals. When edge modifications are required, it ranks candidate edges by combining link prediction scores with their contributions to cycles, and employs a language model to arbitrate cases where structural signals for an edge are in conflict.

Critical strengths

S1.This paper distinguishes between learning positive edge weights on a fixed edge set and "rewiring"and establishes the conditions under which the former fails to eliminate the residual cycles of a fixed prediction flow through Theorem 1.

S2. Experimental design is rational that two operations have same candidates and base models, the thresholds which are determined based on instances with independent random perturbations are applied directly to both types of attacks.

Critical weaknesses

W1. The connection between the theoretical certificate and the classification objective remains insufficiently established. Theorem 1 bounds the residual when fitting a fixed prediction flow with node potentials on a fixed support (§4.2, Eqs. 5–7). In contrast, REWEIGHT jointly trains edge weights and a fresh GCN (§5.1), allowing the predictions defining that flow to change. Successful classification does not require an exact potential fit to the initial flow. Although the partial correlation of 0.38 in Table 1 provides empirical support for the diagnostic, it remains unclear when the certified fitting limitation translates into a classification disadvantage for reweighting. An analysis of cases where the diagnostic succeeds or fails would help clarify this connection.

W2. The interpretation of cycle residuals as evidence that rewiring is needed requires stronger justification. Section 4.1 constructs the comparison matrix, with classes ordered by training frequency, but does not explain why this ordering should reveal structural errors. Proposition 1 characterizes when a comparison matrix can generate circulation; it does not establish that the resulting circulation identifies undesirable graph structure. The binary-class case, where the diagnostic is identically zero regardless of corruption (§4.3), further illustrates its dependence on the class comparison construction. The paper should justify the chosen ordering and establish how alternative comparison matrices affect diagnostic values and routing decisions.

Questions for authors

Q1. Under what conditions does the inability to fit the frozen prediction flow imply a classification disadvantage for REWEIGHT, given that edge weights and a fresh GCN are jointly trained? Q2. Are there instances where the initial cycle residual is large but reweighting still outperforms rewiring? What distinguishes these cases from those where the diagnostic selects the better operation? Q3. What motivates ordering classes by training frequency when constructing KK, and why should this comparison rule reveal structural errors relevant to graph repair? Q4. How sensitive are the diagnostic values and routing decisions to alternative comparison matrices or small changes in class frequencies? Do these changes materially affect routing regret?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9pnm

Thank you for the detailed questions. We clarify the theoretical-to-empirical connection and the role of the comparison matrix below. References are to the proposed revised manuscript.

W1 / Q1: What connects the certificate to classification? The theorem and routing experiment answer different parts of the question. Theorem 1 characterizes what positive weights can achieve when fitting a specified flow on a fixed support. Its two-sided bound provides a quantitative obstruction to that fit. The classifier is then retrained, and may change its flow; we therefore do not use the theorem as a classification impossibility result.

Instead, the frozen-flow residual is a diagnostic available before either repair is trained. Its relationship to the actual operation gap is tested on 370 instances: partial Spearman correlation is 0.38 with 95% CI [0.21,0.52], controlling for corruption rate, density, cycle-space dimension, and dataset (Table 1). The calibrated gate matches or exceeds the better fixed policy on each dataset and incurs only 0.09 Macro-F1 points of average oracle regret (Table 2). These results supply the empirical connection to classification, while keeping the theorem's scope precise. Table 5 separately verifies the fixed-flow weighted-fitting bound.

W1 / Q2: Where is the diagnostic less informative? The weakest tested association occurs under betweenness deletion on Cora and CiteSeer: correlations are 0.29 and 0.26, neither significant after Holm correction, and the mean Rewire–Reweight gaps are negative (Table 7). At 15% deletion on Cora, Always-Reweight achieves 82.24 versus 79.74 for Always-Rewire (Table 10). This contrasts with Metattack, where the correlations are 0.55/0.60/0.52 across Cora/CiteSeer/PubMed and all are significant after correction.

This distinction is consistent with what the residual measures: missing connectivity need not produce the circulation associated with inconsistent cross-class connections. Since the diagnostic uses an augmented candidate support, this is an interpretation of the observed strata, not a theorem that deletion must lower the diagnostic. The stratum-level results characterize a weak regime; they should not be mistaken for a count of individual high-residual misroutes. PRBCD provides a useful counterpoint: on Cora/CiteSeer the gate selects the preferred Reweight operation on 58/60 instances.

W2 / Q3: Why this comparison construction? Proposition 1 identifies the relevant requirement: K must not be a difference of class scores, which would make every induced flow a gradient. The sign comparison under a class order is a simple construction satisfying this requirement for C>=3. Frequency ordering makes the convention depend on training data rather than arbitrary label names.

For one-hot predictions, a cycle involving at most two classes has zero circulation; nonzero circulation requires at least three predicted classes. Figure 2 makes this concrete with a path indexed 0–5 and classes 0,0,1,1,2,2: adding edge (2,4) leaves rho=0, whereas adding (1,4) gives rho=1/12. This explains how a cross-class shortcut can create a detectable inconsistency. It does not label every multi-class cycle as erroneous: the operational relevance is established by the corruption and routing experiments.

W2 / Q4: How much does the choice of K matter? There are three useful distinctions.

  1. For C=3, the non-gradient class-comparison space is one-dimensional. On fixed predictions and support, admissible K matrices therefore yield proportional cycle components up to sign. Squared cycle-contribution rankings are unchanged. This statement concerns the ranking, not invariance of the normalized cycle fraction, whose total-flow denominator can change.
  2. Frequency changes that preserve class ranks leave K, diagnostics, and decisions unchanged. Rank swaps and frequency ties require separate treatment; the invariance argument is conditional on a consistently defined order.
  3. The direct empirical sensitivity test (Table 16, Metattack 15%) changes average accuracy by -0.15 points for prototype-cosine K, -0.26 for native-label order, and -0.93 for random antisymmetric K. Thus, the tested alternatives retain performance within one point of the default. These are downstream accuracy results, not an unreported claim of identical routing decisions or regret.

Binary case. Every 2x2 antisymmetric matrix is a difference of class scores, so circulation cannot distinguish corruption within this comparison family when C=2. The gate explicitly defaults to Reweight. On PolBlogs, average Macro-F1 over the twelve tested cells is 86.16 for Always-Reweight versus 86.08 for Always-Rewire (Table 10). The default is therefore competitive on the evaluated binary benchmark, while the structural limitation remains explicit.

Thank you for your response. The response addressed some of my concerns, and I keep my score.

We sincerely hope the reviewer can clearly point out any concerns that remain unaddressed, which we will carefully address. We also hope our responses and clarifications can help the reviewer further recognize the value of this work. We thank the reviewer once again for the valuable time and the comments.

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

The paper studies how to choose between positive edge reweighting on a fixed candidate support and support editing for noisy or incomplete graphs. A frozen node classifier induces an antisymmetric bilinear edge flow. Its Hodge decomposition yields a cycle residual, and Theorem 1 shows that positive diagonal reweighting cannot make a non-gradient fixed flow exactly representable by node potentials; a spread-dependent bound controls its normalized weighted residual. The Hodge Router thresholds a density-normalized residual, combines link scores with edge cycle contributions to rank edits, and optionally uses a budgeted language-model arbiter. Evaluation covers random corruption, Metattack, and PRBCD poisoning transferred from a surrogate GCN, with routing controls, ablations, and additional diagnostic and cost studies.

Critical strengths
  1. The paper provides a clear mathematical analysis of fixed-flow potential fitting and identifies conditions under which the proposed diagnostic applies. This offers a useful theoretical perspective on graph repair.
  2. The experiments compare different graph repair strategies under multiple attacks. The results reveal differences in the effectiveness of reweighting and rewiring, while the ablation studies help explain the contributions of individual components.
  3. The language model is used only for rewiring after structural selection. This design separates structural and semantic decisions, limits the use of the language model, and makes its contribution and computational cost easier to evaluate.
Critical weaknesses
  1. The theoretical analysis focuses on fixed-flow potential fitting, while the actual reweighting method optimizes a different objective. As a result, the diagnostic does not directly establish when reweighting is insufficient or rewiring is necessary. The connection between the theoretical result and the repair strategy needs stronger justification.

  2. Some reported metrics and edit budgets appear inconsistent, raising concerns about the reliability of the comparisons. In addition, the improvements over competing methods are relatively small in some settings, and further statistical analysis is needed to establish their significance.

  3. The repair process relies on the true perturbation rate to determine the edit budget and uses dataset-specific threshold calibration. These assumptions may not hold in real-world settings, limiting the method's applicability to unseen graphs with unknown corruption levels.

Questions for authors
  1. Can the authors clarify the inconsistencies in the reported recovery metrics and edit counts? Do all compared methods follow the same editing constraints, and how are differences between baselines handled?
  2. Given the relatively small improvements over some competing strategies, can the authors provide paired confidence intervals and statistical tests to support the claimed advantages of the diagnostic and routing strategy?
  3. How would the repair budget be determined when the true perturbation rate is unknown? Can the proposed strategy maintain its effectiveness using a budget estimated from observable information?
  4. How are the dataset-specific routing thresholds selected? Can the authors clarify whether these thresholds generalize to unseen graphs without additional calibration?
2 comments
AuthorsOpenReview ↗

Response to Reviewer dDcc

Thank you for the careful review. References below are to the proposed revised manuscript.

W1: Connecting the certificate to the repair decision. Theorem 1 establishes that positive reweighting cannot exactly fit a non-gradient flow on a fixed support, and bounds its normalized fitting error by rho_0/gamma and gamma*rho_0. It does not equate potential fitting with classification loss. Retraining can change the flow, so we use the frozen flow as a pre-repair diagnostic rather than a certificate about the retrained classifier.

The classification link is tested directly. Across 370 instances, kappa_diag has partial Spearman correlation 0.38 with the Rewire–Reweight gap, with 95% CI [0.21, 0.52] after controlling for rate, density, cycle-space dimension, and dataset (Table 1). More importantly for the decision, the calibrated gate has average oracle regret of 0.09 Macro-F1 points and matches or exceeds the better fixed policy on every dataset (Table 2). The theory motivates the diagnostic; policy comparisons establish its practical value.

W2 / Q1: Budget protocol and recovery metrics. The common budget applies to edge-ranking methods: GPS, ComFy, LLM4RGNN, and our rewiring branch make q deletions and q additions. Adjacency-output baselines retain their published settings; truncating their learned adjacency to q edits would change those methods. A router instance sent to Reweight makes no discrete edits. Sections 5.1 and Appendix D now distinguish these cases.

Table 14 reports actual edit counts. For a ranking method with 2q edits against q perturbations, precision=recall/2. The corrected router precision is 0.39/0.38/0.42 on Cora/CiteSeer/PubMed, consistent with its reported recall and rounding. For adjacency-output methods, precision, recall, and edit count are averaged separately over runs; a ratio of column means need not equal the mean per-run ratio.

At half budget, the router scores 76.32/73.08/85.87; at twice budget, 77.60/74.02/86.95. At twice budget it remains above the strongest competing baseline on each graph (76.4/73.6/86.8). This is a budget-sensitivity comparison, rather than a claim of exactly matched edit counts for every adjacency-output baseline.

W2 / Q2: Statistical support. The Wilcoxon signed-rank tests are two-sided and paired by evaluation seed. The mean differences and “cells ahead” columns summarize performance across dataset/perturbation/rate conditions; these descriptive counts are distinct from the seed-paired test. The within-condition significance counts use Welch tests over ten runs with Holm correction.

Under random corruption (Table 12), gate versus Always-Reweight gives +0.33 points (26/36 cells ahead); gate versus Always-Rewire gives +0.47 (30/36); router versus gate gives +0.34 (36/36); and router versus the better fixed policy of each cell gives +0.36 (35/36). All four seed-paired Wilcoxon p-values are <0.001.

Table 13 reports the following attack results. Differences are accuracy points; “significant” counts wins after the within-condition Welch test and Holm correction.

AttackRouter versusDifferenceCells aheadSignificantWilcoxon p
MetattackJDR+1.5812/128<0.001
MetattackPro-GNN+2.7011/1290.001
MetattackAlways-Rewire+0.969/1270.006
MetattackAlways-Reweight+8.1612/1212<0.001
PRBCDJDR+2.5612/126<0.001
PRBCDPro-GNN+2.9412/127<0.001
PRBCDAlways-Rewire+2.0210/1240.002
PRBCDAlways-Reweight+1.235/1220.41

The router's advantages over JDR and Pro-GNN are significant under both attacks, and it is significantly worse in none of the tested cells. We do not claim significance for PRBCD versus Always-Reweight (p=0.41). Table 1 separately reports bootstrap confidence intervals for the diagnostic correlations, including [0.21,0.52] for kappa_diag. These are correlation intervals, not confidence intervals for paired policy differences.

W3 / Q3: Unknown corruption rate. The gate does not read the corruption rate; only the rewiring branch requires an edit budget. Table 21 isolates budget misspecification under Metattack 15%: doubling the budget costs 0.54/0.53/0.37 points, compared with 1.82/1.47/1.45 when halving it. This demonstrates budget tolerance in the tested regime, not an estimator of the unknown rate.

W3 / Q4: Threshold transfer. The threshold is selected from a fixed 21-level quantile grid on disjoint tuning seeds and reused unchanged under both attacks. Table 9 further transfers a quantile level selected from the other three datasets: average Macro-F1 is 74.83, versus 74.80 for Always-Reweight and 74.32 for Always-Rewire. Target-specific tuning gives 75.07. Thus, the transferable object is the quantile level, not a universal numerical threshold, and the observed cost of transfer is modest.

Thank you for your response. The response addressed some of my concerns, and I keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.