acceptodds
Under review as a conference paper at ICLR 2027

Decodability Is Not Transferability: Uncovering Limitations in Graph Attribute Auditing via Orthogonal Alignment and Invertible Rotation Controls

Abstract

Attribute inference auditing evaluates whether sensitive attributes beyond the intended task can be decoded from model outputs or internal representations - an important privacy concern for large language models that process graphs. Existing audits use output confidence or train attribute classifiers on shadow-model representations, assessing attribute decodability through transfer to target representations. However, whether cross-model transferability reliably reflects target-side attribute decodability remains unclear. We systematically investigate this relationship through orthogonal alignment and invertible rotation techniques. Orthogonal alignment preserves attribute information while adjusting shadow representations’ coordinate directions to better match target representations. Under identical preprocessing, alignment improves transfer performance by up to approximately 0.30 macro AUROC, suggesting that poor transfer may partly arise from coordinate mismatch rather than difficulty in decoding target attributes. To further verify this, we use invertible rotation to change only the coordinate directions of target representations without removing information, reducing the shadow-trained classifier’s transfer performance by approximately 0.16 macro AUROC while leaving the performance of a linear classifier retrained on target representations nearly unchanged. Overall, attribute decodability and cross-model transferability are not equivalent: low transfer scores do not directly imply that target attributes are difficult to decode. These findings reveal the limitations of assessing attribute exposure solely through transfer performance. Transfer-based audits should therefore interpret results in light of the classifier’s training source and the characteristics of the representation spaces.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Attribute Inference, Large Language Models for Graphs, Attribute Decodability, Cross-Model Transfer, Orthogonal Alignment, Invertible Rotation

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_m57XOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether cross-model transferability reliably reflects target-side attribute decodability through orthogonal alignment and invertible rotation techniques. Orthogonal alignment preserves attribute information while adjusting shadow representations’ coordinate directions to better match target representations. Under identical preprocessing, alignment improves transfer performance, suggesting that poor transfer may partly arise from coordinate mismatch rather than difficulty in decoding target attributes.

Critical strengths
  • The paper evaluates the gap between target-side decodability and cross-model transfer in graph–language models.
  • This results show weak cross-model transfer can coexist with strong target-side attribute decodability.
Critical weaknesses
  • Experiments are limited to two models (SI and Llama). It is interesting to include more recent models.
Questions for authors
  • Can authors use other models in the experiments?
5 comments
AuthorsOpenReview ↗

Additional Validation on GLM-4-9B: Decodability–Transfer Gap and Alignment Gains

Thank you for suggesting that we evaluate our findings on additional models. We had also been considering this extension and had begun preparing the experiments. Your suggestion encouraged us to complete the evaluation on GLM-4-9B, testing whether both the decodability–transfer gap and the improvement from alignment also occur on another model. The results show that GLM-4-9B also exhibits a decodability–transfer gap and that alignment improves transfer, further supporting our main findings.

The experiments now cover three model families—S1, Llama, and GLM—within the graph-encoder–projector–language-decoder design studied in the paper. On GLM-4-9B, we independently trained the graph encoder and projector and evaluated the Frozen setting on NCI1 and BACE. Results are averaged over five model seeds, with five classifier initializations per model seed for MLP.

Here, T and S denote representations from the target and shadow models. The arrow means “probe training source → probe test source”: T→T measures decoding on held-out target representations, S→T measures transfer from shadow to target, and S→S measures decoding on held-out shadow representations.

1. Decoding and direct transfer

We compare Raw representations (standardized, without dimensionality reduction) with PCA-64 representations (standardized and reduced to 64 dimensions). For each path, the preprocessing is learned from the probe training data and then applied to its validation and test data:

RepresentationDatasetProbeT→TS→TS→S
RawNCI1LR0.7060.5290.715
RawNCI1MLP0.7100.5300.722
RawBACELR0.9170.5700.921
RawBACEMLP0.9240.5820.926
PCA-64NCI1LR0.6840.5250.684
PCA-64NCI1MLP0.6870.5340.722
PCA-64BACELR0.8970.5970.902
PCA-64BACEMLP0.9170.5850.917

In both spaces, target-trained probes perform better than probes transferred from the shadow model. For example, the Raw MLP reaches 0.710 on NCI1 and 0.924 on BACE for T→T, compared with 0.530 and 0.582 for S→T. The decoding–transfer gap therefore also appears on GLM.

2. Transfer after alignment

For alignment, we use a shared PCA-64 space built from paired validation representations (247 pairs for NCI1; 91 for BACE). We then compare S→T with and without Procrustes alignment under this same preprocessing. The probe is trained on shadow representations and evaluated on target representations in both cases.

DatasetProbeS→T unalignedS→T alignedGain [95% CI]Holm p
NCI1LR0.5080.597+0.089 [0.070, 0.104]0.0026
NCI1MLP0.5300.614+0.083 [0.052, 0.116]0.0231
BACELR0.6090.771+0.163 [0.091, 0.243]0.0231
BACEMLP0.5760.825+0.249 [0.218, 0.280]0.0006

Alignment improves mean transfer in all four comparisons, by 0.083–0.249 AUROC. All four pass Holm correction across this four-test family. The intervals use 10,000 paired bootstrap resamples of the five model seeds. Gains are calculated before rounding.

We also repeated the LR random-rotation controls on both datasets and all five model seeds, keeping PCA-64 fixed. Rotating the target training and test representations together leaves mean T→T AUROC unchanged. Applying the same rotation to both models changes mean S→T AUROC by less than 0.000002. On BACE, rotating only the target reduces S→T from 0.609 to 0.417, while independent rotations give 0.493. Relative to the common-rotation control, the changes and pointwise 95% intervals are −0.191 [−0.268, −0.114] and −0.115 [−0.200, −0.065], respectively. On NCI1, the corresponding changes are smaller and their pointwise intervals include zero. These controls show that changing coordinate directions can change transfer while preserving target-side linear decoding.

These experiments extend the findings to GLM-4-9B in the Frozen setting: target-trained probes decode attributes more accurately than transferred probes, alignment improves transfer, and random rotations can change transfer while preserving target-side linear decoding. We will include these experiments in the manuscript after completing the rebuttal experiments, together with detailed results for each of the five model seeds in the appendix.

Thank you again for this constructive suggestion. It helped us test our findings on another model and strengthen the evidence for distinguishing decodability from transfer.

AuthorsOpenReview ↗

Respectful Reply to the Reviewer: Revision Response and Request for Feedback

Dear Reviewer,

Thank you for reviewing our paper and sharing your feedback. We have addressed your comments in the discussion and provided additional experimental results. If you have a chance, we would appreciate hearing your thoughts on our response. Your feedback would help us identify what needs further clarification and present our findings more clearly in the revised manuscript.

If any questions remain, we would be happy to clarify. Thank you again for your time and consideration.

Area_Chair_WaFAOpenReview ↗

Please respond to the authors' rebuttal

Hello,

The discussion period ends soon. Please respond to the authors' last comments and note whether they changed your evaluation.

Thanks,

Your AC

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

The paper compares target-trained attribute probes with shadow-trained probes transferred to graph-language representations. Target-side decoding remains strong even when direct transfer is near chance. Procrustes alignment improves transfer, while invertible rotations change transfer without reducing linear decodability. Experiments use two molecular datasets, two language models, and additional training and split controls to motivate separate reporting of these audit outcomes.

Critical strengths
  1. The distinction between a failed transferred probe and low target-side decodability is useful for interpreting privacy audits. The alignment and rotation controls make this distinction concrete.
  2. The study includes multiple seeds, matched-training and structure-based split controls, and input-side baselines. It also appropriately avoids interpreting inconclusive LoRA effects as privacy improvements.
Critical weaknesses
  1. For the MLP results, the alignment gain is not separated into the effects of rotation and mean matching. Equation 4 includes both, but no mean-matching-only control is reported. This would help clarify how much of the recovery comes specifically from correcting coordinate directions.
  2. The experiments cover two molecular datasets and a single graph-token classification interface. The authors acknowledge these limits, but the evidence still covers a fairly small part of the graph-language auditing settings used to motivate the study.
  3. The decoding-transfer gap is checked under structure-based splits, but the alignment gain is not (Appendix F). This leaves the alignment result less well supported under stricter data separation.
Questions for authors
  1. For the MLP, what transfer AUROC is obtained by setting Q = I in Equation 4 (mean matching only), compared with full Procrustes alignment?
  2. Does the alignment gain persist when the paired calibration and test graphs follow the structure-based split protocol?
5 comments
AuthorsOpenReview ↗

Alignment Ablations and Generalization Across Structural Groups

Thank you for your careful review and helpful suggestions. Your questions helped us make our claim more precise: we needed to distinguish the benefit of matching the means from the benefit of adjusting coordinate directions, and to test alignment when calibration and test graphs are separated by structure. We have completed these experiments and report the results below. We will include these results in the paper.

  1. How does mean matching alone compare with full alignment?

We tested the requested Q=IQ=I condition. This shifts the shadow representations to match the target mean without changing their directions. Full Procrustes alignment also adjusts the directions through an orthogonal map. The conditions use the same preprocessing, target test data, classifier seeds, and training and selection procedures.

We compare all four conditions below. Mean matching shifts the center; the orthogonal-only condition adjusts directions around the shadow mean; full alignment combines both. The orthogonal map permits reflections as well as rotations. The table shows MLP macro AUROC. Each value averages five model seeds, with five classifier initializations averaged within each model seed.

Dataset/modelTrainingUnalignedMean onlyOrthogonal onlyFull alignment
NCI1/S1Frozen.52.57.61.63
NCI1/S1LoRA.51.52.57.59
NCI1/LlamaFrozen.56.59.58.65
NCI1/LlamaLoRA.50.54.52.62
BACE/S1Frozen.52.53.79.78
BACE/S1LoRA.55.51.77.765
BACE/LlamaFrozen.53.48.81.83
BACE/LlamaLoRA.56.58.76.80

The orthogonal-only condition improves mean AUROC over unaligned transfer in all eight settings, whereas mean matching alone has mixed effects. Direction adjustment therefore gives the more consistent benefit when applied alone. Combining the two steps can produce a larger improvement: for NCI1/Llama/LoRA, AUROC is 0.504 without alignment, 0.542 with mean matching, 0.520 with the orthogonal map alone, and 0.618 with full alignment. This pattern suggests that the two steps can work together; their gains should not be treated as separate, additive contributions. All eight pointwise 95% bootstrap intervals for full alignment minus mean matching are above zero. In the full set of 48 mechanism comparisons covering both LR and MLP, three remain significant after Holm correction for multiple comparisons. We will report all four conditions in the manuscript.

  1. Does alignment still help under structure-based splits?

We tested alignment using the paper's structure-based splits: molecular scaffold groups for BACE and WL structural families for NCI1. Calibration and test graphs come from different groups. We formed calibration pairs by passing the target-validation graphs through both saved models: 328–329 graphs for NCI1 and 121 for BACE. Preprocessing and alignment use these paired representations, without attribute labels or target-test data.

The following table shows MLP results, averaged in the same way as above.

Dataset/modelTrainingUnalignedMean onlyOrthogonal onlyFull alignment
NCI1/S1Frozen.52.51.63.61
NCI1/S1LoRA.51.50.598.603
NCI1/LlamaFrozen.53.54.587.59
NCI1/LlamaLoRA.492.485.585.586
BACE/S1Frozen.62.60.78.82
BACE/S1LoRA.56.61.81.84
BACE/LlamaFrozen.55.59.74.77
BACE/LlamaLoRA.44.43.75.79

Full alignment and the orthogonal-only condition improve mean transfer AUROC in all eight settings, even though calibration and test graphs belong to different structural groups. One plausible explanation is that different molecular families still share structural features, such as graph size and connectivity. If the two models encode these features along systematically different directions, a mapping learned from calibration graphs can also help the probe read them in unseen groups. The results are consistent with this explanation: part of the coordinate mismatch appears to extend across structural groups, allowing alignment to help beyond the groups used for calibration.

We will include these results and the calibration procedure in the manuscript. We will also state the calibration-set sizes and training settings so readers can interpret the two experiments in context.

  1. Coverage of datasets and interfaces

We agree that two molecular datasets and one graph-token classification interface cover only a limited set of graph-language settings. We will state more clearly that our empirical conclusions apply to the settings tested here. Other graph domains, multiple graph tokens, and generative tasks remain important directions for future work.

Together, these experiments show that direction adjustment gives a more consistent benefit than mean matching alone, while combining the two can further improve transfer. This pattern is also observed under the tested structure-based splits. Thank you again for these constructive questions. They helped us explain more clearly what our results show and where further evidence is needed.

AuthorsOpenReview ↗

Respectful Reply to the Reviewer: Revision Response and Request for Feedback

Dear Reviewer,

Thank you for reviewing our paper and sharing your feedback. We have addressed your comments in the discussion and provided additional experimental results. If you have a chance, we would appreciate hearing your thoughts on our response. Your feedback would help us identify what needs further clarification and present our findings more clearly in the revised manuscript.

If any questions remain, we would be happy to clarify. Thank you again for your time and consideration.

Area_Chair_WaFAOpenReview ↗

Please respond to the authors' rebuttal

Hello,

The discussion period ends soon. Please respond to the authors' last comments and note whether they changed your evaluation.

Thanks,

Your AC

Reviewer_usWLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper’s main contribution is comparing target-trained and shadow-trained classifiers for auditing graph representations, in the setting of molecular graphs with a GLM. The goal is to try to determine whether properties (i.e. graph size and density) can be predicted from their internal representations. A classifier that’s trained on the target model measures what can be decoded when it can see the labeled target representations whereas a “shadow-trained” classifier tests whether you can transfer the predictions of a classifier trained on a separate model. On BACE, target-trained classifiers reach 0.93 macro AUROC while classifiers trained on the shadow model and transferred reach 0.449-0.581. But when the authors align the target and shadow representations they find that transfer improves by up to about 0.3 AUROC, and show that rotations can impact transfer while preserving the target’s linear decodability. Overall, I think this is a useful concrete example of why a failed transfer attack doesn’t mean that an attribute is private but I’m not convinced the result goes far enough beyond already well known alignment issues.

Critical strengths
  1. Their main comparison is well controlled as the target-trained + transferred classifiers use the same target test graphs. Alignment also improves MLP transfer in all eight settings which provides good support for coordinate mismatch as the cause.
  2. Multiple seeds, 2 decoder models, and matched-training and structure-based splits. Also disclosure that an untrained GNN also predicts them well which limits their informativeness.
Critical weaknesses
  1. Their main insight seems limited by how closely it follows from the setup i.e. a classifier trained in one coordinate system may not work in another. Though it’s systematic, the paper already discusses prior alignment and representation-evaluation work so I’d like a clearer picture of what someone learns beyond knowing that transfer may require alignment.
  2. The practical privacy implication isn’t yet clear. There’s a large calibration budget which the authors disclose but also the setup for alignment includes 247 NCI1 or 91 BACE paired validation graphs, and target-side decoding requires labeled target representations.
  3. These two molecular datasets have correlated attributes because the other four structural attributes predict a given attribute with 0.993 macro AUROC on NCI1 and 0.999 on BACE in App. B. So I’m uncertain whether this result extends to sensitive attributes or other graph-language models.
Questions for authors
  1. What practical decision would change based on this result, beyond accounting representation mismatch when using a “shadow classifier”?
  2. Do you foresee some “auditor” having access to both the paired representations and labeled target data required here? How does alignment performance change with fewer available paired examples?
6 comments
AuthorsOpenReview ↗

Audit Decisions, Data Access, and Calibration Budget for Graph-Language Attribute Auditing

Thank you for your suggestions on auditing applications and calibration budgets, which helped clarify practical decisions and required data.

Contribution and audit decisions

Beyond existing alignment work, we add controlled evidence in graph-language attribute auditing: using T→T as a direct reference to test whether S→T scores reflect target-side decodability.Under identical preprocessing and on the same target test set, we measure how much transfer alignment recovers, then use invertible rotations to test whether transfer scores can change while target-side linear decodability remains constant. This turns a known coordinate issue into an audit question: does lower S→T support concluding that attributes are harder to recover?

The results affect how audits are interpreted and continued: to assess an external attacker's capability, report S→T under their access conditions; to assess attribute recoverability, evaluate T→T when labeled target samples are available. With paired representations, compare transfer before and after alignment to test whether coordinate mismatch contributes to the low score. Reliable prediction after alignment shows recoverability; low performance after alignment cannot establish that the attribute is hidden. This provides a testable basis for the inference from “an attack failed” to “an attribute is protected.”

Data access

Model owners or authorized evaluators with target representations and labels can conduct T→T; aligned S→T needs labeled shadow data and unlabeled paired representations. For example, auditors train shadow models and probes on public graphs and labels; providers run separate private graphs through the target and auditor's fixed shadow model, returning only paired vectors. Training, calibration, and test graphs are disjoint. Auditors have no arbitrary target-query access, private original graphs, or calibration labels. Evaluators retain test labels and score fixed probes, reducing graph and label sharing while supporting external auditing.

EvaluationPurposeTraining and calibration data
T→TDirectly test target-side decodabilityTarget training representations with attribute labels
Aligned S→TTest transfer after paired calibrationLabeled shadow training representations and paired calibration representations without attribute labels

Calibration budget

We added 16/32/64-pair and full-budget comparisons across NCI1/BACE, S1/Llama, Frozen/LoRA, and LR/MLP. At each budget, only selected pairs fit standardization, PCA-8, and alignment; both conditions share preprocessing and the target test set, with shadow training/validation data unchanged. Macro AUROC averages MLP initializations and three small-budget subset draws within model seeds, then across five seeds and 16 combinations.

Calibration pairsUnaligned S→TAligned S→TMean gainPositive combinations
160.5210.611+0.09015/16
320.5200.616+0.09615/16
640.5200.624+0.10415/16
Full: 247(NCI1) / 91(BACE)0.5340.634+0.10015/16

With 16 calibration pairs, the mean gain is 0.090, positive in 15/16 combinations, showing that improvements also occur at smaller budgets.Increasing the budget gives no consistent monotonic improvement, and recovery varies across settings. With five model seeds, only one of 64 paired comparisons passes Holm correction. We will therefore report effects, pointwise intervals, and per-seed results, without treating similar means as equivalence between small and full budgets.

One possible explanation is that probes learn attributes from shadow training data, while a few pairs calibrate coordinate correspondence. Budgets also change calibration subsets and PCA directions, so gains need not increase monotonically with sample size.

Scope

Appendix B reports attribute correlations and a GNN's performance without task training. We agree that the five attributes do not provide independent evidence and coverage remains limited. However, easy prediction or correlation does not imply that confidentiality is unnecessary: recovering input information that the recipient is not authorized to know remains relevant to auditing. We compare T→T and S→T for the same attribute, without requiring independence between attributes. Thus, low S→T coexisting with high T→T in the tested settings still shows that low transfer alone cannot establish that target attributes have been hidden.

We also independently trained GLM-4-9B's graph encoder and projector and evaluated five model seeds. MLP T→T on NCI1/BACE is 0.710/0.924, versus direct S→T of 0.530/0.582. In shared PCA-64, the four LR/MLP alignment gains are 0.083–0.249, all passing Holm correction. This extends the evidence to a third model family; sensitive attributes and other graph-language interfaces are directions for future validation.

We will add these experiments and access conditions to the paper, with detailed results and uncertainty in the appendix. Thank you again for your constructive suggestions.

AuthorsOpenReview ↗

Respectful Reply to the Reviewer: Revision Response and Request for Feedback

Dear Reviewer,

Thank you for reviewing our paper and sharing your feedback. We have addressed your comments in the discussion and provided additional experimental results. If you have a chance, we would appreciate hearing your thoughts on our response. Your feedback would help us identify what needs further clarification and present our findings more clearly in the revised manuscript.

If any questions remain, we would be happy to clarify. Thank you again for your time and consideration.

Dear Authors, Thank you for the additional experiments. I believe the smaller calibration budgets and additional controls resolve my practical questions. However, I am still concerned that this contribution is not beyond that which has already been published on representation mismatch, and given the datasets and attributes studied. I will keeping my rating at 2. Thank you for your efforts.

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

Paper examines shadow-model attribute auditing for GLMs. Authors compare a classifier trained on target model's representation (T->T) with one trained on shadow model applied to target (S->T). Setup is a GraphSAGE feeling into 8B models. NCI1 and BACE, 5 structural attributes, 7 extraction points, logistic-regresion and MLP probes with 5 model seeds.

Transfer is near-chance while target side decodability is 0.911-0.931. Orthogonal alignment recorders up to about 0.3 AUROC, leaving 0.08-0.21 gap. Consistent rotations leave linear decodability unchanged.

Authors recommend reporting T->T and S->T separately.

Critical strengths
  • Strong experiment design tracking the main claim in the paper
    • Clean comparisons on T->T, S->T and S->S.
    • Authors are clear about attributes being predictable; and demonstrate that: removing task-label effect with scores being high
  • Disclosures are honest: healine gap is descriptive, not formal. 5 seeds should be treated as repetition. Buggy MLP disclosed
Critical weaknesses
  • Separately trained networks are unlikely to line up. This is not a novel or surprising find.
  • Threat model undermines paper's own recommendation: alignment for target's representation come directly from the input graph; so auditor has access -> can compute the labels and train T->T; which explains "mark it unassessed".
  • Attributes aren't sensitive: untrained GNN already decodes them, authors also concede. They are highly correlated
Questions for authors

None

7 comments
AuthorsOpenReview ↗

Direct Target Training, Access Conditions, and the Scope of Attribute Auditing

Thank you for your careful review and recognition of the experimental design and transparent reporting. Your comments helped us clarify the auditing scenarios and threat model for alignment.

Direct target training and access conditions for alignment

We agree: an auditor with the original calibration graphs and their target representations can compute our structural labels and train T→T. The paper already reports T→T using the full target training set; we also completed a comparison with the same calibration-graph budget. Using 247/91 target graphs for NCI1/BACE and averaging over two models, two probe types, and five model seeds, direct T→T has higher mean AUROC in the shared PCA-8 space in all four dataset-by-Frozen/LoRA groups, exceeding aligned S→T by 0.022–0.079; retaining the original dimensions gives differences of 0.066–0.152. Both routes use the same target graphs, while the shadow probe's original training data remain unchanged. We will provide detailed results and uncertainty estimates across budgets and model seeds in the appendix. These results support including direct target training when target attribute labels are available.

Alignment experiments primarily study the decoding–transfer gap, and also apply to auditing with only paired representations shared:

StageData providerExternal auditor
TrainingNo private original graphs need to be sharedTrains the shadow model and probe using public graphs and labels
CalibrationRuns a separate set of private graphs through the target model and the auditor's fixed shadow model; returns only paired vectorsFits alignment from sample correspondence, without calibration graphs or true labels
TestingProvides held-out target representations; retains true labels and scores predictions after the probe is fixedSubmits predictions before and after alignment

The three graph sets are disjoint; the auditor cannot arbitrarily query the target model, and vectors carry no identifiers linking them to original graphs. Thus, the auditor can fit alignment but cannot directly train supervised T→T on these calibration data. This procedure reduces direct sharing of graphs and labels while supporting external auditing; whether the vectors expose attributes still requires testing. The provider can conduct T→T internally as a reference.

Thank you for this suggestion. We will clarify: S→T results should specify unaligned direct transfer or aligned transfer, and the calibration sample size. T→T alone requires no “unaligned” label. Separately, “target-side decodability unassessed” means that T→T has not been evaluated; the two labels have different meanings. Our paper has evaluated T→T.

Furthermore, for attributes whose labels cannot be obtained directly from the original inputs and that are claimed to be hidden, even calibration inputs may not provide attribute labels. An auditor with labeled shadow samples and paired calibration representations can use alignment to improve the shadow probe's predictions on target representations and conduct aligned S→T auditing. Our structural-attribute experiments demonstrate this test's principle; its effectiveness on these other attributes requires further validation.

Significance of the experiments

We agree with your observation that independent networks may not align. This phenomenon need not itself be counterintuitive; what requires testing is its effect on audit judgments. At BACE's final layer, T→T macro AUROC is 0.911–0.931 and direct S→T is 0.449–0.581, across models, training settings, and probes. Under the same preprocessing and test set, alignment recovers part of the transfer performance; invertible rotations change transfer while preserving target-side linear decodability. Thus, “an attribute becoming harder to recover” and “a particular probe becoming harder to transfer” need to be distinguished. These controls can help auditors determine what a score decrease supports and when further testing is needed.

Role and scope of the attributes

Section 4.1 already describes these as input-derived benchmark attributes; Appendix B.3 discusses their strong correlations and prediction by a GNN without task training. We agree they are not five independent disclosures. However, ease of prediction differs from the need for confidentiality: recovering an attribute that the recipient is not authorized to know remains relevant to auditing, even if the information existed before task training. Correlation limits coverage, while controls on the same attribute and test set can still examine whether transfer scores reflect target-side decodability.

After completing the rebuttal experiments, we will incorporate the matched-budget results and access conditions, with detailed results and uncertainty estimates in the appendix. Thank you again for helping us explain the experiments' purposes and the basis for our conclusions more precisely.

Thank you for the detailed clarification and additional experiments.

The distinction between access to original calibration graphs and access only to paired representations addresses an important part of my concern. In the latter setting, the external auditor can fit an alignment without having the ground-truth labels needed for direct supervised T→T training. I therefore no longer regard alignment as necessarily redundant under those restricted access conditions.

For the graph-access setting, the reported matched-budget results support including direct T→T as a baseline. Please include the detailed results and uncertainty estimates in the revision, with sufficient breakdown across models, probe types, seeds, and calibration budgets to assess the consistency of the reported advantage. Also distinguish clearly between procedures evaluated experimentally and the proposed external-auditing workflow.

Appreciate the clarification that the intended contribution is the interpretation of audit scores, rather than the observation that independently trained representations may be misaligned. This narrower framing is useful, but my concern about the degree of novelty remains.

Overall, the response reduces my concern about the access assumptions. I would like the revised paper to make these distinctions explicit and keep its broader auditing claims within the demonstrated scope.

AuthorsOpenReview ↗

Response to Reviewer Feedback and Planned Manuscript Revisions

Thank you for taking the time to consider our clarification and additional experiments. We appreciate your guidance and understand your remaining concern about novelty. Once the discussions with all reviewers have concluded, we will revise the manuscript during the rebuttal phase to include the detailed matched-budget results and uncertainty estimates across models, probe types, seeds, and calibration budgets. We will also clearly distinguish the procedures evaluated in our experiments from the proposed external-auditing workflow and limit our claims to what the experiments support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.