acceptodds
Under review as a conference paper at ICLR 2027

PROBE: Automated Spurious Cue Discovery and Cross-Architecture Shortcut Scoring

Abstract

Image classifiers often rely on shortcuts, yet systematically discovering and comparing such dependencies across classifiers remains challenging. We introduce PROBE, a fully automated two-stage auditing pipeline requiring no predefined attributes, group labels, or retraining. The first stage, Discover, names the visual cues a classifier relies on for each class and tests each cue through a targeted counterfactual edit, measuring the resulting change in classifier confidence to separate object-intrinsic from spurious cues. The second stage, Score, turns the discovered cues into rendered probe images that show the object with its context, the context alone, the object alone, or the object under stacked confuser cues, and reports object-only accuracy, shortcut-stress accuracy, and confidence shift for each classifier. Across five ImageNet classifiers evaluated on a shared candidate set, 86% of shortcut cues that pass each classifier's own confidence threshold are flagged by at least two classifiers and 31% by all five. Probes that show the context alone rarely recover the label after filtering detected render leakage. The discovered catalog recovers $4% of human-annotated reference cues and achieves higher measured per-phrase precision against Salient ImageNet labels than three annotation-free baselines. Debiasing a classifier moves the audit's scores in the expected direction, while an untrained classifier yields no substantive findings, indicating that the scores respond to classifier behavior rather than being determined solely by the tools that construct the probes. The pipeline scales to thousand-class and open-vocabulary settings, where the extreme pairwise differences on the accuracy-based axes reach statistical significance.

Primary area:
interpretability and explainable AI
Keywords:
spurious correlations, shortcut learning, model auditing, interpretability, counterfactual image editing, diffusion models, vision-language models, concept discovery, robustness evaluation, image classification

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

The authors propose PROBE, a training-free framework for automatically discovering spurious visual cues and evaluating shortcut reliance across image classifiers. The framework consists of DISCOVER, which uses VLM-generated cue hypotheses and counterfactual image editing to assess classifier confidence changes, and SCORE, which generates synthetic probes to compare shortcut reliance across different architectures.

Critical strengths

1. Important and Practical Problem.

The paper addresses the important problem of automatically identifying spurious cues learned by image classifiers and comparing shortcut reliance across different architectures, without requiring manually annotated bias attributes.

2. Broad Experimental Coverage and Useful Supplementary Analyses.

The authors evaluate PROBE across multiple datasets and classifier architectures, including CNNs, vision transformers, and vision-language models. They also provide supplementary analyses of VLM and editor sensitivity, synthetic probe fidelity, and the relationship between confidence shift and real-image robustness. These analyses are valuable for understanding both the capabilities and limitations of the proposed framework.

Critical weaknesses

1. Insufficient validity of counterfactual cue attribution.

The DISCOVER stage identifies spurious cues based on classifier confidence changes after individual counterfactual edits. However, the method does not sufficiently verify whether the edits preserve object identity and other relevant visual attributes. Consequently, confidence changes may result from unintended modifications or editing artifacts rather than the targeted cue. Furthermore, evaluating each cue independently may overlook interacting or redundant shortcuts, as removing one cue may have little effect when other shortcuts remain available.

This is a fundamental concern because the reliability of the discovered spurious-cue inventory is central to the proposed framework. Without stronger evidence that the confidence changes reflect the intended cue interventions, the resulting bias assignments remain difficult to trust.

2. Insufficient Validation of the Proposed Shortcut Scoring Metrics.

The SCORE stage is a central contribution of the paper, but it remains unclear whether the proposed metrics reliably measure shortcut reliance. The authors report that confidence shift has a weak association with real-image group robustness. Moreover, a VLM-based fidelity audit (Section 4.1 and Appendix G.12) identifies errors in the synthetic probes: the target object remains present in 41/400 (approximately 10%) bias-stripped images and is absent from 102/381 (approximately 27%) real-cue-only images.

Although the authors provide post-hoc fidelity filtering experiments, it is unclear how much these generation errors affect the reported scores and classifier rankings. This is a critical concern because the paper's main contribution is to compare shortcut reliance across architectures, yet the results may partly reflect image-generation errors rather than actual differences in shortcut reliance.

Questions for authors

1. Novelty relative to prior work.

How does DISCOVER differ from MDC [1] and SaMyNa [2] ? In particular, MDC already combines attribute hypothesis generation, counterfactual image editing, and hypothesis testing for model diagnosis. Could the authors clarify DISCOVER's specific methodological contributions and compare its bias discovery performance with these approaches?

2. Reliability of counterfactual interventions.

How do the authors distinguish confidence changes caused by the intended cue modification from those caused by unintended changes or editing artifacts? Do DISCOVER's cue assignments remain reliable when using independently verified edits?

3. Multiple interacting shortcuts.

How does DISCOVER handle interacting or redundant shortcuts? Could its per-cue confidence threshold miss important shortcuts when removing one has little effect but removing several causes a substantial prediction change? A controlled multi-shortcut benchmark such as UrbanCars [3] could help clarify this issue.

4. Reliability of SCORE rankings.

How do the reported synthetic probe generation errors affect the SCORE results and classifier rankings? In particular, do the main conclusions remain consistent after excluding invalid images, as investigated in Appendix G.12?

[1] Chen, X., et al. (2025). Model Diagnosis and Correction via Linguistic and Implicit Attribute Editing. CVPR.

[2] Ciranni, M., et al. (2025). Say My Name: a Model's Bias Discovery Framework. TMLR.

[3] Li, Z., et al. (2023). A Whac-a-Mole Dilemma: Shortcuts Come in Multiples Where Mitigating One Amplifies Others. CVPR.

5 comments

Response to Reviewer HiUr (1/2)

W1.1 — Reliability of counterfactual edits

DISCOVER uses automatic checks to reduce unreliable edits: a near-identity gate rejects ineffective edits, and the VLM examines the original and edited images when assigning categories. While individual edits are not human-verified, we validate the resulting cue inventories against independent annotations. DISCOVER recovers 84% of Salient ImageNet's reference cues, correctly places the documented spurious cues of all 10 applicable Hard ImageNet classes in the spurious category, and identifies labeled contexts for 55 of 60 NICO++ categories (Sec. 4.2).

We additionally test whether the decisions depend on the image editor. Replacing FLUX with InstructPix2Pix yields 93% agreement (126/135) on cues verified by both editors (App. G.17). Together, these results provide complementary evidence for the reliability of DISCOVER's outputs, beyond any individual editing model.

W1.2 — Testing each cue alone may miss interacting or redundant shortcuts

We agree that testing cues individually may miss dependencies that only emerge when multiple shortcuts are removed together. To investigate this, we added a joint-removal experiment in which all discovered contextual cues are removed simultaneously for each of the 100 classes (new table is added to App. G.18).

Across five classifiers, joint removal reveals only 0–3 additional classes per classifier in which the confidence change exceeds the threshold but no individual cue removal does. Thus, in nearly all cases where joint removal reveals a dependency, at least one individual cue already produces a measurable effect.

Furthermore, SCORE evaluates combinations of contextual cues rather than isolated cues, allowing it to measure their combined influence on classifier predictions. These results suggest that individual cue testing captures most dependencies detectable through joint removal in our experiments.

W2.1 — Reliability of SCORE's shortcut-reliance metrics

SCORE measures classifiers' responses to discovered visual cues, rather than serving as a direct measure of real-image robustness. We provide three complementary experiments supporting its sensitivity to classifier behavior.

First, on Waterbirds, we evaluate the same ResNet-50 backbone before and after two debiasing interventions (DFR and GroupDRO), using identical probes. Both interventions improve real-image worst-group accuracy (from 0.751 to 0.908 and 0.858, respectively), while SCORE's confidence shift and shortcut-stress accuracy move in the expected directions (Sec. 4.3).

Second, an untrained ResNet-50 produces no substantive cue findings and near-chance probe accuracy, indicating that the generative pipeline alone does not produce the observed results.

Third, our additional evaluation on real Spurious ImageNet photographs shows an association with SCORE confidence shift in the expected direction, providing complementary evidence from a real-image benchmark (new App. G.19).

Together, these experiments support SCORE's ability to detect shortcut-related changes in classifier behavior. We clarify in Sec. 4.3 that confidence shift is primarily interpreted within model families, rather than as a calibrated measure of robustness across architectures.

W2.2 — Generation errors in SCORE probes

We acknowledge these fidelity issues and explicitly report the VLM grader's findings: the target is flagged as present in 10% of context-only probes and absent in 27% of object-only probes (App. G.12). Target leakage can inflate context-only accuracy, which we therefore interpret as an upper bound on genuine context-only recognition. Conversely, missing targets can underestimate object-only accuracy.

We also evaluate the impact of these errors directly. Fidelity filtering substantially reduces context-only accuracy, while preserving the classifier pairs defining the extremes of all three SCORE rankings when the evaluated classes are held fixed (App. Table 25 and new Table 26). We discuss this controlled comparison further in Q4.

Response to Reviewer HiUr (2/2)

W2.3 — Effect of generation errors on scores and rankings

As detailed in Q4, we added a controlled comparison that isolates the effect of fidelity filtering by holding the evaluated 78 classes fixed. Removing flagged renders changes the measured scores but preserves the classifier pairs defining the extremes of all three SCORE rankings (App. Table 25 and new Table 26).

Q1 — Comparison with MDC and SaMyNa

DISCOVER differs from MDC and SaMyNa in three key respects: it does not require model errors, tests individual cues through counterfactual confidence changes, and explicitly distinguishes object-intrinsic from spurious cues.

SaMyNa extracts ranked keywords from image captions without counterfactual testing or separating object-intrinsic from spurious features. MDC goes further by identifying attributes shared by misclassified examples and testing their influence through image editing and accuracy changes. However, its discovery is restricted to model errors, and it does not explicitly separate intrinsic from spurious attributes.

In contrast, DISCOVER can identify dependencies even when a classifier predicts correctly. It proposes open-vocabulary cues, tests each through targeted edits, and categorizes them using the classifier's confidence response. Furthermore, these inventories feed into SCORE, enabling shared cross-architecture evaluation rather than stopping at cue discovery (Sec. 2–3).

Q2.1 — Separating intended cue effects from editing artifacts

While we do not independently verify every edit, two complementary checks support the reliability of DISCOVER's outputs. Replacing FLUX with InstructPix2Pix yields 93% agreement on cues verified by both editors (App. G.17). Furthermore, the resulting cue inventories agree with independent reference annotations across Salient ImageNet, Hard ImageNet, and NICO++. Together, these results provide evidence that DISCOVER's findings are not solely driven by artifacts of a particular image editor.

Q2.2 — Reliability of cue assignments on independently verified edits

Our validation focuses on the resulting cue inventories rather than independently verifying individual edits. We evaluate these outputs against reference annotations from Salient ImageNet, Hard ImageNet, and NICO++, finding strong agreement in cue recovery and the identification of documented spurious features. These provide independent evidence for the reliability of DISCOVER's outputs.

Q3.1 — Can the per-cue threshold miss interacting or redundant shortcuts?

As detailed in W1.2, we added a joint-removal experiment to test this concern. Across 100 classes, only 0–3 classes per classifier exhibit a threshold-crossing confidence change under joint removal when no individual cue removal does. This suggests that single-cue testing captures most dependencies detectable through joint removal in our experiments.

Q3.2 — Evaluation on UrbanCars

UrbanCars is a domain-specific benchmark using artificially composited images with deliberately planted shortcuts (backgrounds and co-occurring objects). Reliably manipulating these shortcuts requires access to the compositing elements, which our general-purpose VLM and diffusion editor are not designed for. Evaluating UrbanCars would therefore require domain-specific editing or re-compositing. Our new joint-removal experiment (W1.2) directly examines interacting shortcuts in PROBE's intended natural-image setting. We clarify this scope in App. G.1.

Q4 — Do SCORE's conclusions hold after fidelity filtering?

Yes. To isolate the effect of fidelity filtering, we added a new controlled comparison using the same 78 classes as the filtered evaluation in Table 25. We re-scored these classes without filtering, so the only difference between the two evaluations is the removal of flagged renders.

Filtering preserves the classifier pairs defining the extremes of all three SCORE rankings. The confidence-shift and shortcut-stress gaps remain essentially unchanged, while the object-only accuracy gap increases after removing renders with missing target objects. The differences between the original 99-class evaluation and the 78-class evaluation also reflect changes in class composition, which this controlled comparison eliminates.

The new results are reported in App. Table 26.

Thank you for the detailed rebuttal. However, my main concerns remain partially unresolved.

DISCOVER reliability: Given the low Jaccard similarity (0.128) between editors, how stable are the discovered spurious-cue inventories? The 93.3% agreement is limited to cues verified by both editors.

SCORE validity: While I understand that SCORE is not intended to directly measure real-image robustness, the near-zero correlation (−0.05) on Waterbirds raises questions about its reliability for cross-architecture comparisons. Could the authors further clarify what supports such comparisons?

Reviewer_y5y4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PROBE, an automated auditing pipeline that discovers visual shortcuts in image classifiers and compares their behavior without pre-defined attributes, group annotations, or retraining. PROBE is composed of two stages: (1) DISCOVER uses a vision-language model to propose visual cues and editing instructions and applies diffusion-based counterfactual edits to categorize each cues and (2) SCORE then converts the discovered cues into shared synthetic probes and evaluates classifiers along different diagnostic axes.

Critical strengths

Automated, open-vocabulary cue discovery grounded in measured model responses. Rather than stopping at saliency maps or plausible textual explanations, PROBE turns candidate cues into testable hypotheses and records the classifier’s response to each intervention. It reduces the manual effort needed to investigate previously unspecified shortcuts and provides empirical evidence for subsequent inspection.

Critical weaknesses
  • Insufficient validation that DISCOVER’s category assignments reflect genuine classifier reliance. In the ImageNet-100 audit, 78.6% of classifier-edit pairs receive VLM judgments and pool-level gating also permits substantive assignments even when the individual classifier’s confidence change is below the threshold. The rules-only analysis is informative, but it substantially changes the cue inventories and does not establish that edits isolate the intended cue without altering object identity or other predictive information. Similarly, agreement with Salient ImageNet validates correspondence with saliency-visible cues, not the correctness of category judgments.
  • SCORE remains confounded by image generation. The paper’s VLM grader flags missing targets in approximately 27% of object-only probes and unintended targets in approximately 10% of context-only probes. Although the post-hoc filtering is a useful check, the grader is not validated against human labels (possibly remaining some undetected errors) and the ImageNet-LT and LVIS probes are not fidelity-filtered. Evaluating all classifiers on identical images does not eliminate these confounds because models may respond differently to generation artifacts. ConfShift shows little association with real-image group robustness and cross-family confidence scales complicate its interpretation. Although the authors acknowledge these limitations, they leave unresolved whether SCORE measures the shortcut reliance identified by DISCOVER or sensitivity to the generated evaluation distribution.
Questions for authors
  • Can the validity of DISCOVER’s interventions be independently verified? Could the authors provide a small, randomly sampled audit of the archived edits, assessed independently of the VLM’s verdict and divided by rule-based versus VLM-assigned decisions?
  • Do SCORE’s main conclusions survive fidelity filtering when the evaluated classes are held fixed? The probe-fidelity audit in Table 25 is useful, but it uses 78 classes, whereas Table 1 uses 99. The reported comparison changes both image quality and class composition.
4 comments

Response to Reviewer y5y4 (1/2)

W1 - Do DISCOVER's category assignments reflect genuine classifier reliance?

We agree that distinguishing cue discovery from the correctness of category assignments is important. We clarify the role of the reported percentages and provide complementary validation of DISCOVER's outputs.

First, the 78.6% VLM-assignment rate and pool-level gating are properties of the decision procedure, not measures of its accuracy. The VLM assigns categories when deterministic rules do not apply, while pool-level gating allows a shared edit to proceed to categorization when at least one classifier exceeds the confidence threshold. These mechanisms determine how the cue inventory is constructed; they should not be interpreted as evidence of correctness.

Second, we evaluate the resulting cue inventory against independent human annotations. DISCOVER recovers 529 of 630 (84%) Salient ImageNet reference cues using semantic matching, with the similarity threshold selected through held-out validation (App. G.4). This evaluates the complete discovery pipeline, including its gating and assignment procedures, rather than individual components. However, as the reviewer correctly observes, this experiment establishes agreement on cue identity, not the correctness of category assignments.

Third, we provide additional evidence for the spurious-category assignments using datasets with documented cues. On Hard ImageNet, DISCOVER assigns the explicitly documented spurious cue to the spurious category for all 10 classes with named reference cues. On NICO++, its spurious inventory identifies a labeled context for 55 of 60 categories, for both ERM and GroupDRO classifiers (Sec. 4.2). Unlike Salient ImageNet's category-agnostic cue recovery, these comparisons directly assess whether known spurious cues appear in the intended category.

Together, these evaluations provide complementary evidence for DISCOVER's reliability: Salient ImageNet validates open-vocabulary cue recovery, while Hard ImageNet and NICO++ verify that independently documented spurious cues are assigned to the intended category. Furthermore, the untrained-classifier control produces no substantive findings (Sec. 4.3), confirming that cue assignments depend on the audited classifier's responses rather than VLM-generated descriptions alone.

W2.1 - Reliability of the VLM-based fidelity check

We agree that the VLM grader has not been independently validated against human annotations. However, the grader is used only for post-hoc fidelity analysis, not to produce the original SCORE results. Our main concern is therefore whether its filtering leads to meaningful changes in the evaluation.

To address this concern, we added a new random-filter control experiment (App. G.12). On the same 78 classes, we compare the original unfiltered results, VLM-based fidelity filtering, and random filtering that removes the same number of images per class and probe variant, repeated over 200 seeds.

The results show that VLM filtering reduces context-only accuracy for every classifier, whereas random filtering fails to reproduce this reduction in any of the 200 trials. This provides evidence that the grader selectively removes images contributing to inflated context-only predictions, rather than simply improving results by reducing the evaluation set.

Importantly, the main SCORE conclusions remain consistent with and without fidelity filtering, including the confidence-shift ranking. Thus, the new control supports the usefulness of the fidelity check while showing that our conclusions do not rely on applying it. The revised manuscript reports the experiment in App. G.12 and clarifies the grader's role and limitations in Sec. 4.1.

W2.2 - Fidelity filtering of ImageNet-LT and LVIS probes

We agree that ImageNet-LT and LVIS have not undergone the same fidelity filtering as ImageNet-100. These larger-scale experiments primarily demonstrate PROBE's scalability, while ImageNet-100 provides a more thoroughly validated evaluation. Importantly, fidelity filtering on ImageNet-100 preserves the confidence-shift ranking (new results in App. G.12).

Extending the fidelity audit to ImageNet-LT and LVIS requires grading and potentially regenerating probes across approximately 1,100 additional classes. While this is beyond the current rebuttal period, we commit to completing this analysis and including the results in the final version.

W2.3 - Sensitivity to generation artifacts across classifiers

We agree that different classifiers may respond differently to generation artifacts, even when evaluated on identical images. SCORE partially mitigates this concern by measuring confidence shift within each classifier, comparing its responses to different probe variants rather than directly comparing confidence values across architectures. However, variant-specific artifacts may still affect these differences.

Response to Reviewer y5y4 (2/2)

W2.3 (continued)

To assess this empirically, we re-scored the same 78 classes after removing renders flagged by the fidelity grader (new results in App. G.12). The confidence-shift ranking remained unchanged, suggesting that the observed ordering is not driven solely by the detected rendering errors. We have clarified the interpretation of confidence shift and the potential influence of generation artifacts in Sec. 4.3.

W2.4 — Interpretation of confidence shift and cross-family comparisons

Confidence shift is a diagnostic of a classifier's response to contextual cues, not a proxy for real-image group robustness. It measures the change in target-class confidence between bias-heavy and object-only probes. Because confidence scales differ across architectures, we interpret its magnitude primarily within the same model family, while cross-family comparisons rely more directly on the accuracy-based axes.

Our Waterbirds debiasing experiment supports this interpretation (Sec. 4.3). Using the same ResNet-50 backbone and identical probes, both DFR and GroupDRO reduce the magnitude of confidence shift while improving real-image worst-group accuracy (from 0.751 to 0.908 and 0.858, respectively). This demonstrates that the metric responds in the expected direction to interventions that reduce shortcut reliance.

The new Spurious ImageNet evaluation also shows an association in the expected direction, although the correspondence across five classifiers is limited. We have clarified the intended interpretation and limitations of confidence shift in Sec. 4.3.

W2.5 — Does SCORE measure shortcut reliance or sensitivity to generated images?

We provide two complementary checks to distinguish shortcut reliance from generation artifacts. First, in our Waterbirds experiment, we evaluate the same ResNet-50 backbone before and after two debiasing interventions (DFR and GroupDRO), using identical probes. By holding the generated images fixed, we isolate changes in classifier behavior. Both interventions improve real-image worst-group accuracy, while SCORE's confidence shift and shortcut-stress accuracy move in the expected directions (Sec. 4.3).

Second, our fidelity-filtering analysis shows that removing renders flagged as problematic preserves the confidence-shift ranking (new results in App. G.12), indicating that the observed ordering is not driven solely by these detected rendering errors. Together, these experiments provide evidence that SCORE is sensitive to changes in shortcut reliance, rather than merely reflecting properties of the generated images. We clarify in Sec. 4.3 that cross-architecture differences may nevertheless also reflect model-specific sensitivity to synthetic imagery.

Q1 — Can counterfactual edits be verified independently?

We agree that independent verification of the counterfactual edits would provide a valuable additional assessment of DISCOVER. Our current validation focuses on the resulting cue inventories rather than human annotation of individual edits.

Specifically, the complete DISCOVER pipeline, including both rule-based and VLM-assigned decisions, recovers 84% of the human-annotated reference cues in Salient ImageNet (Sec. 4.2). We also report a rules-only analysis alongside the full pipeline (App. G.13), allowing the effect of VLM-based category assignment to be examined separately. Complementary evaluations on Hard ImageNet and NICO++ further assess whether documented spurious cues appear in the appropriate category.

These experiments provide independent evidence for the quality of the resulting cue inventories, although they do not directly assess the fidelity of individual edits. We clarified this distinction in the revised manuscript (Sec. 4.3).

Q2 — Do the SCORE conclusions hold when the evaluated classes are held fixed?

Yes. To address this concern, we added a new controlled comparison that isolates the effect of fidelity filtering (App. Table 26). We re-evaluated the original unfiltered probes on exactly the same 78 classes used in the fidelity-filtered analysis (Table 25), ensuring that the only difference between these evaluations is the filtering.

The results show that fidelity filtering does not change the classifier pairs defining the extremes of any of the three SCORE rankings. The confidence-shift and shortcut-stress gaps remain virtually unchanged, while the object-only accuracy gap increases after removing renders in which the target object was missing.

Thus, the differences between the original 99-class results and the 78-class analysis primarily reflect the change in class composition, rather than fidelity filtering itself. The new comparison is reported in App. Table 26.

I appreciate the authors for additional experiments with clarifications. The fixed class comparison and random filter experiment address several of my concerns.

However, my main concern about whether DISCOVER’s category assignments are supported by valid interventions is not sufficiently answered. The rules-only analysis and untrained classifier experiment are useful sanity checks, but neither establishes the semantic validity of the interventions.

Thank you for revisiting our response. We are glad that the additional experiments have addressed several of your concerns.

We agree that the rules-only and untrained-classifier experiments are sanity checks. However, three other results provide complementary evidence supporting DISCOVER's category assignments:

  1. Agreement with documented shortcuts. On Hard ImageNet, DISCOVER assigns the documented spurious cues to the correct category for all 10 applicable classes. On NICO++, it identifies a labeled context in the spurious category for 55 of 60 classes, for both ERM and GroupDRO classifiers (Sec. 4.2). These results directly evaluate category assignments against independent references.

  2. Consistency across editors. Replacing FLUX with the architecturally different InstructPix2Pix yields 93.3% agreement (126/135) on cues verified by both editors (App. G.17). This suggests that the shared assignments are not specific to one editor's artifacts.

  3. Independent real-image evidence. DISCOVER's cues match the human-annotated spurious features in 11 of 12 overlapping Spurious ImageNet classes (App. G.19), providing corroboration beyond edited images.

Together, these results support the reliability of DISCOVER's outputs across independently annotated datasets and different editing architectures. While they do not establish that every individual edit perfectly isolates the intended cue, we believe they provide substantial evidence for the validity of the resulting category assignments.

Reviewer_m3U2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a way to audit classifiers. The idea is a 2 stage approach: First, a VLM is used to detect candidate cues. A diffusion editor then changes each cue, and the classifier confidence change is measured. Stage 2 makes four types of synthetic probe images for each class. Then, it ranks a pool of classifiers on object only acc, shortcut stress acc and confidence shift. The authors validate results on Imagenet.

Critical strengths

The paper addresses an important problem. Auditing classifiers helps understand what they rely on when making predictions and it helps make it interpretable. The idea of using VLM to detect the spurious cues, using diffusion models to edit and measure score change is a simple approach and makes a lot of sense.

Critical weaknesses

I think the novelty of the approach is really marginal. Each part of PROBE already exists in prior work. What is new is the combination: a per-cue counterfactual test, plus a shared four-variant probe set and a three-axis ranking of many classifiers.

The paper says that no prior work combines open-vocabulary discovery, a per-cue counterfactual test, and shared cross-architecture scoring, which may be true. But each part is known, and the combination is a minimal engineering contribution.

The paper does not cite related work such as ASPIRE (language-guided spurious feature discovery with image editing), Counterfactual Alignment (Cohen et al., TMLR 2025), Spurious ImageNet (Neuhaus et al.), Wiles et al. (2022), ImageNet-D (Zhang et al., CVPR 2024), and ViG-Bias (Marani et al., ECCV 2024).

Questions for authors

The auditing process this way is very time consuming for each sample. Is there a way to make it faster? Moving away from classifiers, when using VLM for doing classification, can you use a similar approach? The experiments are done on a relatively small scale. Are there plans to scale it up?

2 comments

Response to Reviewer m3U2 (1/2)

W1.1 - Novelty beyond combining existing components

PROBE addresses a fundamental limitation of existing shortcut audits: how can we identify and measure a classifier's reliance on visual cues that were not known or specified in advance? Existing approaches often require predefined attributes, known biases, or user-specified interventions. PROBE instead discovers class-specific cues in an open vocabulary, including unexpected shortcuts, without bias annotations, attribute lists, or retraining.

Crucially, PROBE does not treat these discovered cues as explanations simply because a VLM proposes them. Each cue becomes a testable hypothesis: DISCOVER performs a targeted counterfactual edit and measures the audited classifier's confidence change. This distinguishes plausible visual descriptions from cues with measurable influence on the classifier's predictions, and separates object-intrinsic from spurious dependencies.

The contribution extends beyond identifying individual shortcuts. SCORE converts the discovered cue inventory into a shared set of controlled probes, allowing classifiers with different architectures and training procedures to be evaluated on the same visual evidence. It measures three complementary aspects of their behavior: object-only accuracy, shortcut-stress accuracy, and confidence shift. Thus, PROBE connects previously unspecified cue discovery, classifier-specific intervention, and systematic cross-architecture evaluation in a single automated auditing framework.

The effectiveness of this approach is supported empirically. DISCOVER recovers 529/630 (84%) Salient ImageNet reference cues, identifies the documented spurious cues of all 10 Hard ImageNet classes with explicitly named cues, and identifies a labeled context for 55/60 NICO++ categories (Sec. 4.2). It also achieves higher measured per-phrase precision than three annotation-free discovery baselines.

We agree that VLMs, diffusion editors, and synthetic evaluation sets are established technologies. Our contribution is not a new generative component, but an auditing methodology that makes their outputs testable against classifier behavior and reusable for comparative evaluation. The components are replaceable: VLM substitutions preserve the bias-heavy/bias-stripped accuracy drop in a five-class pilot, while an alternative diffusion editor agrees with FLUX on 93.3% of the 135 cues verified by both editors (App. G.17). These experiments support the framework's modularity, while also showing that cue inventories depend on the chosen components.

W2 - Missing related work, Spurious ImageNet results

We thank the reviewer for identifying these relevant papers. We have incorporated all six into the revised manuscript and clarified their relationship to PROBE. The key distinctions concern how visual cues are discovered, how their influence on a classifier is tested, and how the resulting evidence is used for evaluation.

Discovery and counterfactual explanation. ASPIRE identifies and removes spurious features from training images to improve a classifier through retraining; PROBE instead audits an existing classifier without modifying it. Counterfactual Alignment searches for prediction-changing directions in latent space and evaluates them using attribute classifiers, requiring an attribute classifier for each feature. Wiles et al. generate images, identify classifier failures, and describe them through captions, restricting discovery to failures on generated images. ViG-Bias augments error-based slice discovery with Grad-CAM information and requires labeled errors. In contrast, PROBE discovers open-vocabulary cues from class images and classifier attention, tests their influence through targeted edits, and produces class-level inventories without requiring misclassifications, predefined attributes, or retraining.

Model-specific versus predefined evaluation cues. ImageNet-D evaluates robustness using a predefined set of backgrounds, textures, and materials. PROBE's SCORE instead constructs evaluation probes from the cues identified during DISCOVER, enabling evaluation of dependencies discovered for the audited classifier pool. As an additional comparison, we evaluated our five classifiers on ImageNet-D. The background-induced accuracy drop has Spearman correlation 0.90 with SCORE's confidence shift (new App. G.19), although this is not statistically significant with five models.

Response to Reviewer m3U2 (2/2)

W2 (continued)

New experiment: comparison with Spurious ImageNet. Spurious ImageNet evaluates classifiers on real photographs containing predefined spurious features without the target object. While these photographs avoid synthetic generation artifacts, collecting such examples requires substantial curation and restricts evaluation to the documented shortcuts. PROBE instead automatically discovers class-specific cues and constructs corresponding evaluation probes, without requiring a predefined shortcut inventory or manually collected feature-only images. This makes SCORE applicable to new classification settings and enables cross-architecture comparisons using a shared probe set.

We additionally evaluated all five classifiers using Spurious ImageNet's released code (new App. G.19). At the cue level, DISCOVER recovers the documented feature in 11 of 12 overlapping classes, supporting its ability to identify independently documented shortcuts. At the classifier level, Spurious ImageNet AUC and SCORE confidence shift show an association in the expected direction, although the relationship is modest across the five evaluated classifiers. Real feature-only photographs distinguish the classifiers more strongly on this benchmark, highlighting a trade-off between the fidelity of curated real-image evaluation and the generality of automatically generated probes. Full results and analysis are now provided in the new App. G.19 section.

Q1 - Can the per-sample audit be made faster?

PROBE prioritizes automating a task that would otherwise require substantial manual effort: discovering, testing, and categorizing previously unknown visual shortcuts. This involves multiple VLM and diffusion-model calls, making the current implementation computationally demanding. However, many of these operations can be parallelized across images and candidate cues. Moreover, when auditing multiple classifiers, the expensive cue-proposal and image-generation steps are shared across the pool rather than repeated for each classifier (Sec. 3.3). PROBE is also modular, allowing faster VLMs and image editors to be substituted without changing the auditing procedure. Further efficiency improvements are possible, while our primary focus is establishing the feasibility and effectiveness of automated, open-vocabulary shortcut auditing.

Q2 - Can the same approach audit a VLM used as a classifier?

Yes. PROBE only requires a classifier to provide confidence scores over the target labels, making it applicable to vision-language models. Our evaluated pool already includes CLIP ViT-B/32 and SigLIP2-base used as zero-shot classifiers.

To further test this capability, we audited a generative VLM, Qwen2.5-VL-7B, by presenting the target class and five confusing classes as candidate answers. We evaluated it on the same archived counterfactual edits and synthetic probes as the original classifier pool. Its discovered cue counts and confidence shift fall within the range of the other models, and it exhibits the same substantial accuracy drop when contextual cues are removed.

These results demonstrate that PROBE can audit generative VLMs used for classification, not just conventional image classifiers (new App. G.20). Although the VLM's classification task uses fewer candidate labels, making absolute accuracies not directly comparable, the experiment confirms that the auditing procedure transfers to this setting.

Q3 - Are there plans to scale the experiments up?

We have already extended PROBE substantially beyond the 100-class audit. The full pipeline has been evaluated on all 1,000 ImageNet classes using ImageNet-LT and on 1,032 open-vocabulary LVIS categories with zero-shot classifiers. We have also expanded the classifier pool to 13 models by re-scoring the existing counterfactual edits and probes, without requiring additional image generation. This demonstrates an important scalability advantage: once the probe set is constructed, additional classifiers can be evaluated at relatively low computational cost.

We retain the 100-class audit as our primary detailed analysis because its edits and probes are archived and subjected to additional fidelity checks. Together, these experiments demonstrate PROBE's scalability across both class vocabularies and classifier architectures (Sec. 4.1, Apps. G.10 and G.14).

Comments on the paper

1 comment

We thank the reviewers for their constructive feedback. We have revised the manuscript, added five new experiments, and clarified the methodological contributions and limitations of PROBE. Below, we summarize the main changes and address concerns shared across reviews. Detailed responses follow individually.

Main contribution and recognized strengths

We appreciate that the reviewers recognize the importance and practical value of automated shortcut auditing. Reviewers m3U2 and HiUr highlight the relevance of the problem, with m3U2 describing the approach as simple and sensible. Reviewer y5y4 particularly recognizes the value of open-vocabulary cue discovery grounded in measured classifier responses: rather than stopping at saliency maps or plausible textual explanations, PROBE turns candidate cues into testable hypotheses. Reviewer HiUr also highlights the broad experimental coverage and analyses of component sensitivity and probe fidelity.

Building on these strengths, we would like to clarify PROBE's central methodological contribution. PROBE enables systematic auditing of visual shortcuts that are not known or specified in advance. It discovers class-specific cues in an open vocabulary, tests their influence through counterfactual interventions, and converts the resulting inventories into shared probes for cross-architecture evaluation. The contribution is not any individual VLM, diffusion editor, or evaluation metric, but the automated auditing framework that connects these capabilities without requiring predefined attributes, bias annotations, or retraining.

Summary of changes in the revised manuscript

  • Joint removal of context cues. We added a joint-removal experiment that tests whether single-cue edits miss interacting shortcuts (new App. G.18).
  • Spurious ImageNet and ImageNet-D. We evaluated our five classifiers on both benchmarks and compared the results with SCORE (new App. G.19).
  • Generative VLM as the audited classifier. We audited Qwen2.5-VL-7B on the same archived edits and probes as the original classifier pool (new App. G.20).
  • Controlled fidelity comparison. We re-scored the unfiltered probes on the same 78 classes as the filtered evaluation, so that filtering is the only difference (new Table 26 in App. G.12).
  • Random-filter control. We compared the VLM grader's filter with random removal of the same number of renders over 200 seeds (new results in App. G.12).
  • We also expanded the related work and clarified the interpretation and limitations of SCORE in Sec. 4.3.

Reliability of DISCOVER (Reviewers y5y4 and HiUr)

Although individual counterfactual edits are not human-verified, we validate the resulting cue inventories against independent reference annotations. DISCOVER recovers 84% of Salient ImageNet reference cues, assigns the documented spurious cues of all 10 applicable Hard ImageNet classes to the spurious category, and identifies labeled contexts for 55 of 60 NICO++ categories (Sec. 4.2). Furthermore, replacing FLUX with InstructPix2Pix yields 93% agreement on cues verified by both editors (App. G.17). These complementary evaluations support the reliability of DISCOVER's outputs across reference datasets and editing models.

Reliability of SCORE (Reviewers y5y4 and HiUr)

We explicitly report synthetic probe fidelity issues and added two experiments to assess their impact. First, holding the evaluated 78 classes fixed, fidelity filtering preserves the classifier pairs defining the extremes of all three SCORE rankings (new App. Table 26). Second, VLM-based filtering reduces context-only accuracy across all five classifiers, whereas matched random filtering does not reproduce this reduction in any of 200 trials (App. G.12). These results support the effectiveness of fidelity filtering and the stability of the main SCORE conclusions.

Interpretation of confidence shift (Reviewers y5y4 and HiUr)

We also clarify that confidence shift measures a classifier's response to contextual cues, rather than serving as a direct measure of real-image robustness. On Waterbirds, debiasing the same ResNet-50 backbone with DFR or GroupDRO improves real-image worst-group accuracy while moving SCORE's metrics in the expected directions on identical probes (Sec. 4.3). An untrained classifier produces no substantive findings, while the new Spurious ImageNet comparison provides complementary real-image evidence (App. G.19). We have clarified the interpretation of confidence shift, particularly across model families, in Sec. 4.3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.