acceptodds
Under review as a conference paper at ICLR 2027

REND: Reject-Encoder Novelty Diffusion for Out-of-Distribution Detection on Graphs

Abstract

Existing methods for out-of-distribution (OOD) detection on graphs read their score from a representation that was trained for another purpose, namely classifier logits, classifier embeddings, or diffused class evidence. When the observed classes are fewer than the embedding dimensions, every score read from the logits is invariant along a residual subspace of the embedding. Scores that read the whole embedding escape this invariance, but classification training never requires any direction of the representation to separate novel nodes from familiar ones, so whatever novelty structure a borrowed representation retains is incidental. We propose REND (Reject-Encoder Novelty Diffusion), in which a dedicated reject encoder reads the raw features and the graph structure directly and is trained by novelty objectives alone. Since OOD labels are not available at training time, a self-supervised surrogate-outlier loss supplies the missing supervision, and stop-gradients keep the reject branch from reaching the classifier, which leaves the accuracy unchanged. A personalized-PageRank kernel then smooths the learned reject propensity, and its teleport floor guarantees that every node keeps a fixed share of its own novelty. REND shows excellent performance on twelve homophilous, heterophilous, and heterogeneous benchmarks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Network, Out of Distribution Detection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_cnnYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes REND, a graph OOD detection method that learns to identify OOD nodes without relying solely on the classifier's representations. REND consists of: 1) a reject encoder trained using raw features and the graph structure, and 2) a personalized PageRank kernel, which propagates OOD scores across the graph while retaining part of each node's original score. The reject encoder's output is combined with the classifier's energy score to produce the final OOD score after diffusion, without affecting in-distribution node classification. Experiments on twelve graph datasets under leave-out-class shifts demonstrate improvements over existing graph OOD detection methods. I have several concerns regarding the method's motivation and experimental settings. However, I would consider increasing my score if these concerns are adequately addressed.

Critical strengths
  1. REND presents a novel novelty-learning framework for graph OOD detection, separating OOD detection from node classification to capture better OOD detection signals.

  2. The paper provides theoretical analysis of the limitations of logit-based OOD scores and the score retention properties of personalized PageRank diffusion.

  3. The authors have benchmarked against a wide range of baselines, achieving consistent performance gains across diverse datasets.

Critical weaknesses
  1. The paper focuses only on leave-out-class OOD detection, with only a small feature corruption study (Figure4) in the appendix. However, existing works typically evaluate on more types of shifts, such as structural perturbations and combined feature-structure shifts. It is recommended to either explore the performance of REND on more commonly evaluated shift scenarios or clearly discuss why they are not compared.

  2. Additionally, Line 316 states that REND evaluates label-leave-out shifts by withholding the last half of the classes. Although Appendix Figure 4 investigates different held-out class selections, it doesn't evaluate different numbers of held-out classes or analyse the distributional similarity between ID and OOD classes. This is important because the paper explicitly argues that logit-based OOD scores become more limited when fewer classes are observed.

  3. The reject encoder is trained with a shuffled feature (Lines 269-271), which can introduce inconsistencies between node attributes and their neighborhoods. However, OOD nodes may have internally consistent features and structural relationships. The paper has not discussed whether the learned detector captures general distributional novelty or primarily recognizes the artifacts introduced by its training procedure.

  4. Proposition 1 identifies a limitation of logit-based OOD scores, but it does not show how the reject encoder captures novelty information unavailable in the classifier's full representation. More direct representation analysis would be helpful to bridge the connection between the theoretical findings and the proposed design.

  5. Missing more recent graph OOD detection discussions in the related work, such as:

  • Song et al., Breaking One-Size-Fits-All: Revisiting Out-of-Distribution Detection on Graphs Under Diverse Distribution Shifts. AAAI 2026
  • Wang et al., What Information Matters? Graph Out-of-Distribution Detection via Tri-Component Information Decomposition. ICML 2026.
  • Chen et al., Decoupled Graph Energy-based Model for Node Out-of-Distribution Detection on Heterophilic Graphs. ICLR 2025

General Notes: It would be helpful to include some intuitive explanations for each theoretical claim, the current writing can be difficult to understand and follow.

Questions for authors
  1. Why is a completely separate GNN encoder necessary (which is what REND can essentially be viewed as)? Could a simpler auxiliary detector, a shared encoder with an additional OOD objective, or a capacity-matched alternative achieve comparable performance. Although the ablation studies demonstrate its effectiveness, it does not fully establish whether the improvements come from learning more informative novelty representations, additional model capacity, or the separate optimization objective.

  2. What evidence shows that the reject encoder actually learns information unavailable in the classification representation? The null space argument does not seem to full address this. Can the authors directly measure and compare the novelty information retained by the two representations?

  3. Appendix Figure 4c shows that REND performs similarly to softmax under Gaussian corruption, despite showing stronger improvements under Bernoulli corruption. Why and when does the dedicated reject encoder would provide limited benefits (which also connects to weakness 1)?

5 comments

We thank the reviewer for the careful reading. All changes in the revised PDF are marked in blue.


W1. More shift types. We added structure, feature-interpolation, and combined shifts on all twelve graphs in the single-graph protocol of the paper (one tenth of the test nodes perturbed in place; Fig. 8c). Mean AUROC:

ShiftGNNSafeNODESAFERENDREND localFree energy
Structure57.060.163.067.672.0
Feature57.558.358.460.963.7
Both59.262.065.770.575.8

We report this result as it is. The shifted nodes here are isolated, not a connected population, so by Proposition 3 diffusion has nothing to separate, and every propagated score falls below its local counterpart. REND loses less to diffusion than averaging and stays ahead of the two propagation detectors on structure and combined shift, and adding rewiring and interpolation to its surrogate adds two points (65.1, 67.8). The best score on isolated perturbations is the undiffused free energy, which REND's untouched classifier supplies unchanged. The revision states this scope in the main text: the reject channel targets novelty that arrives as a population, such as unseen classes, and the local energy of the same model covers isolated perturbations.

W2. Number of held-out classes and ID/OOD similarity. We varied the number of held-out classes on five graphs, from one class to all but two (22 configurations, Fig. 8d). REND has the highest AUROC in all 22 against free energy, GNNSafe, NODESAFE, and Mahalanobis. Its margin over the best of the four is 4.2 points when the fewest classes are held out and 12.1 when only two classes remain for training (Amazon-Photo 94.1 against 73.9; Wiki 86.2 against 75.5), which is the trend Proposition 1 predicts. For similarity, across the 78 held-out classes of the twelve benchmarks, per-class detection shows no significant association with the cosine similarity between the class's feature centroid and the nearest training class (Spearman 0.20, p = 0.08); the thirds by similarity score 83.0, 80.8, and 85.1 from least to most similar.

W3. Surrogate artifacts versus general novelty. New paragraph "What the Surrogate Teaches". The nodes REND is evaluated on are genuine held-out-class nodes, with consistent features and edges within their class, not shuffled nodes. A linear probe separates them at 86.0 on the reject embedding. A surrogate with labeled-ID donors only still reaches 79.7. Under pure structure shift, where no feature is altered and rewiring was never seen in training, the local reject score still reaches 67.6. The ID anchor rewards any direction along which labeled nodes are tight and corrupted ones are not, and such directions also set apart unseen classes.

W4 / Q2. Direct representation analysis. Fig. 7c. Linear probe for ID versus OOD (diagnostic only, it uses OOD labels): logits 63.6, classifier embedding 75.3, reject embedding 86.0, with the reject embedding ahead on all twelve graphs. Label-free Mahalanobis: 67.6 on the classifier embedding, 78.6 on the reject embedding. Only 4% of the classifier embedding's variance lies in the null space of the head, so classification training leaves those directions nearly empty, as part (ii) of Proposition 1 suggests.

W5. Related work. Song et al. (AAAI 2026), Wang et al. (ICML 2026), and Chen et al. (ICLR 2025) are now cited and discussed in Related Work.

Q1. Is a separate GNN encoder necessary? New controls, all with the same surrogate and losses (Fig. 7b):

Variant (same surrogate and losses)AUROCID accuracy
Head on detached classifier embedding71.981.7
Same, classifier widened to 128 units / deepened to 5 layers69.5 / 63.481.8 / 75.1
Shared encoder (stop-gradient removed)77.670.3
NODESAFE++ with our surrogate as exposure set79.977.3
Graph-free (MLP) reject encoder81.581.9
One-layer reject encoder79.381.7
REND83.081.9

Capacity-matched alternatives (wider or deeper classifier with a novelty head) score 69.5 and 63.4. A shared encoder with the additional OOD objective reaches 77.6 but accuracy falls from 81.7% to 70.3%. A simpler graph-free reject encoder reaches 81.5 and a one-layer one 79.3, against 83.0. Most of the gain comes from a separate representation trained by the novelty objective; reading the graph adds the rest (Facebook: 75.1 graph-free against 88.1).

Q3. Gaussian versus Bernoulli corruption. Scale-matched Gaussian noise lowers the classifier's confidence, which the energy already reports, so a second representation has little to add. Bernoulli vectors produce large activations that the classifier reads as confident, and only a representation trained for novelty flags them. The reject encoder helps most when a shift leaves classifier confidence intact or inflates it, and adds little when confidence already drops and the shifted nodes are isolated, consistent with the new shift study. This explanation is now in the appendix.

General note on intuition. Proposition 1 is now followed by a plain-language sentence with a worked count of the invisible directions, Proposition 2 is read as a stopping random walk and now points to the separation result it supports, and Proposition 3 comes with an interpretation paragraph and a measured value of its factor on every benchmark.

Thank you for the detailed rebuttal and additional experiments, I appreciate the authors' efforts. Some of my original concerns are adequately addressed, however, three critical concerns remain. Thus, i would like to maintain my rating:

  1. The new experiments provide valuable evidence but also shows an important limitation in the W1 Table, where REND underperforms both its local variant and free energy across all three additional shift types. I appreciate the theoretical explanation that diffusion is designed for connected OOD populations. However, this narrows the demonstrated effectiveness of the method primarily to leave-out-class shifts, which affects the practicality of the method.

  2. Moreover, i am still not fully convinced on the response to W4. Particularly, while the linear probe results show better OOD separability in the reject embedding, i do not think this necessarily establish that the reject encoder captures novelty information unavailable in the classifier's full representation, so the connection to Proposition 1 still appears insufficiently established.

  3. Also, the graph-free reject encoder achieves 81.5 AUROC compared with REND's 83.0 in the Q1 response, suggesting that much of the improvement comes from the separate novelty objective rather than graph-based encoding. I believe the architecture complexity is thus not well explained given this performance comparison.

Reviewer_G7i3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes REND, a graph OOD detector that learns a dedicated rejection representation instead of relying entirely on classifier features. The reject encoder is trained using feature-shuffling corruption, and its output is combined with classifier energy before PPR smoothing. Gradient separation preserves the classifier’s training objective. The reported results are strong: REND achieves a mean AUROC of 83.3% across twelve datasets, compared with 73.1% for the strongest baseline.

Critical strengths
  1. The most convincing result is the representation ablation. Replacing the dedicated encoder with a head on detached classifier embeddings costs 11.8 AUROC points on average. The comparison with a shallow reject encoder also helps address whether the improvement simply comes from additional depth.
  2. The gains extend across different graph families, rather than being concentrated on one benchmark. Reporting multiple detection metrics and variability over ten runs strengthens the empirical case. I also appreciate that the appendix examines restricted transductive access and reports substantial failures under ID-only subgraph training. These experiments help establish where the method works and where it does not.
Critical weaknesses
  1. The ablations establish the value of the learned reject representation more clearly than the marginal contribution of PPR diffusion. The detached-embedding experiment provides useful evidence for the reject encoder, but the reject-channel removal simultaneously removes both the learned representation and its propagation, while the energy-only experiment changes the local input score. The paper also observes that PPR and GNNSafe-style averaging perform similarly on average when both operate on classifier energy. What remains missing is a controlled comparison that holds the learned reject score fixed and evaluates direct local scoring, ordinary neighborhood averaging, and PPR diffusion. Such a comparison would clarify whether PPR materially improves the proposed reject representation.

  2. The method depends substantially on transductive access to the deployment graph. In Table 9, training only on the ID-induced subgraph reduces mean AUROC from 83.3% to 65.1%, below GNNSafe’s 72.5%. Label-free adaptation on the full deployment graph recovers the score to 78.5%, but this recovery again requires access to the deployment mixture. This dependence is compatible with the explicitly transductive setting, but it limits claims that the reject encoder learns novelty features that transfer beyond the mixture observed during training. In addition, the restricted-target experiment simultaneously changes the locations and number of corrupted nodes, so it cannot determine whether the degradation is caused by reduced spatial coverage or simply by less surrogate supervision.

  3. Figure 6 and the misclassification-detection claim require clarification. At r=0.1r=0.1, panel (a) appears to place REND below GNNSafe, whereas panel (c) reports a positive mean REND-over-GNNSafe margin. These quantities should agree if they use the same datasets, runs, and aggregation procedure; otherwise, the paper should explain the difference in aggregation. The misclassification-detection result should also be described more cautiously because it uses classifier confidence rather than the learned reject score. It therefore primarily demonstrates that the separated reject branch preserves classifier behavior, rather than showing that the proposed rejection mechanism detects classification errors.

Questions for authors
  1. How much does PPR improve over direct reject scoring and ordinary averaging when the learned input signal is held fixed?
  2. Does restricting corruption to labeled ID nodes still cause a large drop when the corruption budget is matched?
  3. How much of the adaptation gain in Table 9 can be recovered by recalibrating the scores while keeping the reject encoder frozen? This would help clarify whether the degradation mainly reflects a calibration mismatch or a limitation of the learned representation.
  4. Since the reject branch is trained independently of the classifier, how sensitive is detection performance to selecting checkpoints by ID validation accuracy?
2 comments

We thank the reviewer for the careful reading. All changes in the revised PDF are marked in blue.


W1 / Q1. PPR with the learned score held fixed. We now hold the trained reject propensity fixed and change only the operator that spreads it (Fig. 7a):

Readout of the same learned scoreMean AUROC
Local (no propagation)82.3
Lazy averaging, 1 / 2 / 10 / 50 steps82.1 / 81.7 / 78.6 / 73.8
PPR kernel, 2 / 5 / 10 / 50 steps83.0 at every horizon
PPR also lowers mean FPR95 from 56.6 to 55.3, while two-step averaging raises it to 56.9. Trained through each operator: none 80.8, averaging 80.8, PPR 83.0. PPR therefore improves the learned representation modestly on average and substantially on label-mixed graphs (Chameleon +13.4, Wiki +5.5 over averaging), and it removes the depth sensitivity of averaging. We agree that the representation is the main contribution and the revision says so.

W2 / Q2. Transductive dependence and matched corruption budget. We ran the matched-budget control (Fig. 8a). Targets and donors on labeled nodes: 66.5. The same number of corrupted positions spread over all nodes, donors still labeled ID: 79.5. Same budget, unrestricted donors: 83.6, level with the default 83.0. The drop is caused by spatial coverage and not by the amount of surrogate supervision. We also agree with the scope of the claim and have revised the text: the reject encoder learns novelty features from the unlabeled deployment graph, the resource every transductive detector reads, and we do not claim transfer to an unseen graph without adaptation.

Q3. Recalibration with a frozen reject encoder. New control (Fig. 8b), same 100 label-free steps, restricted parameters:

Parameters updatedMean AUROC
None (ID-subgraph training as is)64.0
Readout scalars only (steepness, offset, rate)65.1
Scalars + novelty head, encoder frozen66.4
Full reject branch79.0

Recalibration recovers at most 2.4 of 15.0 points. The degradation is a limitation of the representation learned on the ID-only subgraph, not a calibration mismatch, and the revision states this.

Q4. Checkpoint selection by ID validation accuracy. We recorded detection AUROC at every epoch. REND: selected checkpoint 83.0, last epoch 81.1, best epoch in hindsight 86.3, standard deviation 1.8 over the epochs after selection. Softmax: 71.6, 72.1, 77.2. The rule costs both methods a few points against an oracle no label-free protocol can use, and REND leads by at least nine points under each rule.

W3. Figure 6 and the misclassification claim. Panels (a) and (c) use the same runs and agree: at r = 0.1 REND scores 67.3 against 74.8 for GNNSafe, a mean margin of −7.5 in both panels. The curve above zero at r = 0.1 in panel (c) is the per-dataset line of Chameleon (+3.5); the thick line is the mean. The caption and text now state this, including the negative margin. We also reworded the misclassification result: it uses classifier confidence, so it shows that the separated reject branch preserves classifier behavior, not that the reject mechanism detects classification errors.

Dear Reviewer (G7i3), We would like to sincerely thank you for your valuable feedback and thoughtful consideration. If you feel that your concerns have been satisfactorily addressed, we would be truly grateful if you could kindly consider revising the score upward, as it would greatly improve the chances of our paper being accepted.

Reviewer_3DU8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies out-of-distribution detection for transductive node classification. It proposes REND, which learns a dedicated reject encoder directly from node features and graph structure instead of deriving novelty solely from classifier representations. The reject branch combines its learned novelty signal with detached classifier energy and applies personalized PageRank smoothing. It is trained using labeled in-distribution anchors and synthetic outliers generated by feature shuffling. The paper motivates this design through logit-space invariance and score-retention properties of diffusion. Experiments on twelve benchmarks report substantial detection improvements, approximately unchanged classification accuracy, and extensive component and deployment-setting analyses.

Critical strengths

1.The motivation that OOD scores derived directly from classifier outputs may be insufficient for graph OOD detection is reasonable and well motivated.

2.The empirical evaluation is fairly comprehensive, covering 12 benchmarks with diverse properties and reporting both OOD detection and ID classification performance.

Critical weaknesses

1.The presentation lacks a clear logical structure, making it difficult to identify the paper’s central insight for separating ID and OOD or enlarging the ID/OOD gap. Several arguments, such as the null-space discussion in Proposition 1, seem only loosely connected to the actual method design, while the large number of symbols and engineering details further obscure the core idea.

2.The paper argues that classifier-based scores are insensitive to novelty because the classifier is not trained to distinguish ID from OOD, yet the proposed reject propensity combines novelty with energy, where energy is itself derived from the classifier; this appears somewhat inconsistent with the motivation, and a principled explanation of why the combination should outperform novelty alone would strengthen the paper.

3.The surrogate-OOD construction based on random feature replacement raises two questions: 1) whether the useful OOD signal mainly comes from introducing genuinely OOD-like features or whether replacing features with those from other ID samples is already sufficient, and 2) whether much of the performance gain may come from this surrogate-OOD generation step itself rather than from the more complex components of REND.

4.It is unclear why propagating OOD scores with PPR should make ID and OOD more separable at the global level. Proposition 2 shows that individual nodes can partially retain their own novelty after propagation, but it does not establish that propagation increases the overall ID/OOD separation, which is important given the central role of PPR in the method.

5.Although the experiments cover 12 benchmarks, most baselines are relatively traditional methods, while several more recent approaches—including score-propagation methods conceptually related to PPR—are missing, making it difficult to fully assess the competitiveness and effectiveness of REND against current methods.

Questions for authors

See Weaknesses.

3 comments

We thank the reviewer for the careful reading. All changes in the revised PDF are marked in blue.


W1. Structure, central insight, symbols, and the role of Proposition 1. We revised the presentation on each of these points.

Central insight. The introduction now states it in one sentence: REND widens the ID/OOD gap at its source, by learning the novelty score in a representation that is not confined to the image of the classification head, and uses diffusion only to average out the node-level noise of that score.

Structure and symbols. The opening of Section 3 now gives the whole method as one chain of four quantities, ν_i → a_i → c_i → u_i (novelty score, affinity, acceptance propensity, diffused OOD score), and names the formal result behind each link. A new notation table (Table 10, Appendix A.13) lists every symbol by role and keeps these four apart from the problem inputs, the classifier, the reject branch, the diffusion, and the training settings.

Proposition 1 and the method. Each part fixes one decision, and each decision has its own control. Part (i): no logit score sees the at least d − K embedding directions outside the head (61 of 64 with three training classes), so ν_i is not a function of the logits. With the energy alone as input the mean AUROC is 72.8, against 83.3. Part (ii): the cross-entropy never supervises those directions, and the classifier embedding places only 4% of its variance there, so reading the whole embedding cannot recover them and ν_i comes from a separate encoder. The same novelty head on the classifier embedding scores 71.9, against 83.0 (table under W3). A linear probe agrees: ID versus OOD is decodable at 75.3 from the classifier embedding and at 86.0 from the reject embedding, in that order on all twelve graphs (Fig. 7c).

W2. Combining novelty with classifier energy. Proposition 1 limits what a logit score can express; it does not say that what the energy does express is wrong. We added this explanation after Eq. (1) and report the numbers openly in the appendix: energy alone 72.8, novelty alone 83.1, their sum 83.3. On the leave-out-class benchmarks the novelty term carries the detection, and the energy changes the mean by less than run-to-run spread, with gains on Wiki (+3.0), DBLP (+1.8), and CiteSeer (+1.3) and a loss on DBLP-H (−3.1). We keep it because it costs no parameter and no forward pass, and because under shifts that act on classifier confidence it is the stronger cue (new shift study, Fig. 8c). A novelty-only affinity is an equally valid configuration, and the paper now says so.

W3. Surrogate construction. (1) Features of other ID samples are sufficient. With donors restricted to labeled ID nodes, so that no OOD vector can enter the surrogate, REND reaches 79.7; at a matched corruption budget the same restriction gives 79.5 (Fig. 8a). (2) The surrogate matters but does not explain the gain alone (table below). Without the surrogate REND loses six points; with the identical surrogate, a head on classifier embeddings reaches 71.9, and NODESAFE++ reaches 79.9 while losing 4.4 accuracy points. REND reaches 83.0 with accuracy unchanged.

Variant (same surrogate and losses)AUROCID accuracy
Head on detached classifier embedding71.981.7
Same, classifier widened to 128 units / deepened to 5 layers69.5 / 63.481.8 / 75.1
Shared encoder (stop-gradient removed)77.670.3
NODESAFE++ with our surrogate as exposure set79.977.3
Graph-free (MLP) reject encoder81.581.9
One-layer reject encoder79.381.7
REND83.081.9

W4. Why PPR increases global separation. We added Proposition 3. For input scores with mean gap Δ between the populations and noise variance σ², the probability that a diffused OOD score falls below a diffused ID score is at most 1/(1 + γ²Δ²/(2σ²)), where γ_ik = √2(ρ_i − ρ_k)/‖V_i − V_k‖₂ and ρ is the kernel mass a node places on OOD nodes. Diffusion improves on the local score for a pair exactly when γ_ik > 1, that is, when the OOD node keeps its mass among OOD nodes while the noise is averaged. We measured the factor on all twelve graphs: the median γ lies between 1.08 and 1.80, the share of pairs with γ > 1 between 65% and 99%, with ρ of 0.76 to 0.97 for OOD nodes against 0.01 to 0.20 for ID nodes (Fig. 7d). On Chameleon, lazy averaging has γ > 1 for 27% of pairs and PPR for 65%, which matches the 13-point gap between the two operators there. The empirical counterpart holds the trained reject score fixed and changes only the operator that spreads it (Fig. 7a):

Readout of the same learned scoreMean AUROC
Local (no propagation)82.3
Lazy averaging, 1 / 2 / 10 / 50 steps82.1 / 81.7 / 78.6 / 73.8
PPR kernel, 2 / 5 / 10 / 50 steps83.0 at every horizon

PPR adds 0.7 points as a readout and 2.2 points when the encoder is trained through it. The paper states plainly that the kernel is a small share of the margin and that the learned reject representation is the main source.

W5. Recent and score-propagation baselines. We added GRASP (Ma et al., NeurIPS 2024), a recent score-propagation method, to Table 1, implemented from the official code on the shared GraphSAGE classifier. It is the strongest baseline (mean 78.3) and wins four graphs. REND remains best on eight of twelve graphs and second on three more, at 83.3 mean AUROC, with AUPR 90.8 against 86.7 and FPR95 55.2 against 65.8. The headline margin in the paper is updated from 10.2 to 5.0 points accordingly. The recent detectors of Song et al. (AAAI 2026), Wang et al. (ICML 2026), and Chen et al. (ICLR 2025) are now discussed in Related Work. We also apply GRASP's propagation to REND's learned score in place of PPR: the mean is 82.4, and CiteSeer, PubMed, and DBLP rise to 90.4, 84.6, and 77.4, so the learned score and a stronger propagation operator are complementary.

Dear Reviewer (3DU8), We would like to sincerely thank you for your valuable feedback and thoughtful consideration. If you feel that your concerns have been satisfactorily addressed, we would be truly grateful if you could kindly consider revising the score upward, as it would greatly improve the chances of our paper being accepted.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

We thank all three reviewers. This comment lists the summary of our rebuttal. Revised text is marked in blue in the PDF.

Reviewer 3DU8

ConcernsComments
W1. Unclear structure and central insight; role of Proposition 1Stated the central insight in the introduction, opened Section 3 with the method as one chain of four quantities, and added a notation table (Table 10). Each part of Proposition 1 now fixes one design decision and has its own control.
W2. Combining novelty with classifier energy versus the motivationClarified after Eq. (1) that Proposition 1 limits what a logit score can express, not that the energy is wrong. Reported energy alone 72.8, novelty alone 83.1 and their sum 83.3, and stated that a novelty-only affinity is equally valid.
W3. Whether the surrogate alone explains the gainDonors restricted to labeled ID nodes still reach 79.7 (Fig. 8a). With the identical surrogate, a head on classifier embeddings reaches 71.9 and NODESAFE++ 79.9 at a loss of 4.4 accuracy points, against 83.0 for REND with accuracy unchanged.
W4. Why PPR increases global ID/OOD separationAdded Proposition 3 and measured its factor on all twelve graphs (Fig. 7d). With the learned score held fixed, PPR gives 83.0 against 82.3 for the local score (Fig. 7a). The revision states that the kernel is a small share of the margin.
W5. Missing recent and score-propagation baselinesAdded GRASP (Table 1), the strongest baseline at 78.3. REND remains best on eight of twelve graphs at 83.3, and the headline margin is updated from 10.2 to 5.0 points. Three recent detectors are discussed in related work.

Reviewer G7i3

ConcernsComments
W1, Q1. Contribution of PPR with the learned score held fixedAdded the controlled comparison (Fig. 7a): local 82.3, averaging 82.1 falling to 73.8, PPR 83.0 at every horizon. The gain is modest on average and large on label-mixed graphs (Chameleon +13.4).
W2, Q2. Transductive dependence; matched corruption budgetRan the matched-budget control (Fig. 8a): 66.5 with targets on labeled nodes, 79.5 with the same budget spread over all nodes. The drop comes from spatial coverage. We no longer claim transfer to an unseen graph without adaptation.
W3. Figure 6 panels; misclassification claimPanels (a) and (c) agree, with a mean margin of −7.5 at r = 0.1; the positive curve is the Chameleon line. Corrected the caption and reworded the misclassification result as preserved classifier behavior.
Q3. Recalibration with a frozen reject encoderAdded Fig. 8b. Recalibration recovers at most 2.4 of 15.0 points, so the degradation is a limitation of the representation, not a calibration mismatch.
Q4. Sensitivity to checkpoint selectionREND scores 83.0 at the selected checkpoint, 81.1 at the last epoch and 86.3 at the best epoch, and leads Softmax by at least nine points under each rule.

Reviewer cnnY

ConcernsComments
W1. Only leave-out-class shiftsAdded structure, feature and combined shifts on all twelve graphs (Fig. 8c) and stated the scope: on scattered perturbations the free energy of the same model is best. Read before diffusion, the reject propensity is above GNNSafe, NODESAFE and GRASP under all three shifts (Appendix A.14).
W2. Number of held-out classes; ID/OOD similarityREND is best in all 22 configurations on five graphs (Fig. 8d), and its margin grows from 4.2 to 12.1 points as fewer classes remain. Detection shows no significant association with class similarity (Spearman 0.20, p = 0.08).
W3. Surrogate artifacts versus general noveltyAdded the paragraph "What the Surrogate Teaches". Evaluation nodes are genuine held-out-class nodes, and the local reject score reaches 67.6 under structure shift, which was never seen in training.
W4, Q2. Novelty information in the two representationsCorrected our claim: with OOD labels a whitened probe reads the classifier embedding almost as well (83.9 against 86.5). The information lies in the null space of the head and stays unreadable without labels; the reject encoder makes it readable (Appendix A.14).
W5. Missing recent related workSong et al., Wang et al. and Chen et al. are now cited and discussed.
Q1. Is a separate graph encoder necessary?Added controls with the same surrogate and losses (Fig. 7b). Capacity-matched classifiers score at most 69.5, and a shared encoder loses 11 accuracy points. Most of the gain comes from the separate objective; the graph adds 1.5 points, rising to 3.8 when features are scarce (Fig. 12).
Q3. Gaussian versus Bernoulli corruptionGaussian noise lowers classifier confidence, which the energy already reports; Bernoulli vectors inflate it. Now explained in the appendix.
General note. Intuition for the theoretical claimsAdded plain-language interpretations after Propositions 1 to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.