- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper studies whether contrastive hardness of synthetic corruptions is sufficient to determine the quality of an error auditor trained to distinguish those corruptions from an observed data pool. The central theoretical result shows that it is not. With the observed pool, error prevalence, representation, and latent real-error process held fixed, the authors construct two corruption distributions with identical complete binary testing regions, and therefore identical optimal ROC curves, Bayes risks, and f-divergences, while their population-optimal contrastive scores rank the real errors in opposite directions. The paper further provides a finite-pool characterization of when such reversals are possible.
The second main theoretical contribution gives a positive result: audit regret can be bounded by the discrepancy between the synthetic corruption distribution and the real error distribution, together with the contrastive AUC shortfall. This motivates fidelity to the target error process, rather than hardness alone, as the relevant property of synthetic corruptions.
The theoretical claims are complemented by a broad empirical study across knowledge-graph anomaly detection, image auditing, summarization, information extraction, and question answering. Crossed interventions show that changing the training corruption can change which errors are detected well, and changing only the evaluated error construction can even reverse the ranking of two published detectors on FB15k-237. Additional experiments test matched measured difficulty and human-annotated error processes. The paper also proposes a benchmark diagnostic based on the maximum AUC obtainable from coarse benchmark partitions and shows that simple priors can outperform published detectors in some settings.
The paper makes a strong conceptual point: difficulty of distinguishing synthetic negatives from the data is not, by itself, evidence that the resulting auditor will rank real errors correctly.
Critical strengths
- Strong and unusually sharp theoretical identification result.
The main theorem is substantially stronger than showing that a particular hardness metric is insufficient. The authors hold the latent error process, observed pool, prevalence, and representation fixed, while constructing corruptions with the same complete binary testing region but opposite optimal audit rankings. Since the testing region determines the optimal ROC curve, binary Bayes risks, and f-divergences, this rules out a broad class of hardness-only arguments rather than attacking one particular metric. I view this as the most important contribution of the paper.
- Careful empirical methodology and unusually transparent treatment of negative results.
The empirical section includes crossed corruption/error interventions, a prespecified replication in an independent detector architecture, an image experiment with matched measured contrastive difficulty, and evaluations involving human-identified errors. Importantly, the authors also report when prespecified strong criteria fail, when effects become unresolved after refitting, when transfer fails, and when simpler same-label controls outperform the proposed matching approach. This makes the empirical evidence more credible and supports the broader conclusion that corruption choice can materially affect audit conclusions.
Critical weaknesses
- The positive theoretical result has a substantial operational gap.
Theorem 2 identifies fidelity to the target error distribution as the relevant quantity, but the discrepancy term depends on the unknown real-error distribution. Consequently, the result does not directly provide a label-free procedure for selecting a good synthetic corruption. In practice, determining whether an auxiliary distribution is sufficiently faithful appears to require target-error information, which is precisely what motivates synthetic corruption-based auditing in the first place. Thus, the negative result is strong and actionable as a warning, while the positive prescription is considerably less operational.
- The empirical evidence is broad but heterogeneous and only partially supports a general practical effect.
The experiments illustrate corruption dependence, but they do not directly instantiate the population theorem. Several empirical results are also weaker than the headline theoretical result: the image study matches a measured discrimination proxy rather than the complete testing region; the image and summary studies miss their prespecified strong criteria; the FRANK interaction is largely driven by one error type; CoNLL++ is unresolved; the QA generic comparison becomes unresolved when fitting uncertainty is included; and the unchanged QA heads fail to transfer to Data2txt. These results do not invalidate the central theorem, but they make the empirical evidence for a broadly recurring practical phenomenon less decisive than the theoretical contribution.
Questions for authors
- Theorem 2 depends on a discrepancy between the auxiliary distribution and the unknown real-error distribution. Is there any practically estimable, label-free or low-label surrogate that the authors believe can be used to select among candidate corruptions? If not, could the authors clarify what actionable corruption-selection procedure they recommend when target-error labels are scarce?
This would materially affect my assessment of the practical significance of the positive result.
- Theorem 1 allows unconstrained corruption distributions and the paper explicitly notes that it does not impose locality or semantic-edit constraints. Are there important commonly used corruption families for which the reversal construction can be ruled out, or for which hardness does become informative about audit quality?
A characterization, even for a restricted family, would help determine how directly the impossibility result applies to practical negative-sampling procedures.
- Could the authors clarify how they believe the empirical evidence should be weighted relative to the theoretical result? Several studies produce positive interactions, but several prespecified strong criteria fail, some contrasts become unresolved after accounting for refitting, and the Data2txt transfer experiment fails. Is the intended empirical claim only that corruption dependence can occur in realistic fitted systems, rather than that matching corruption semantics generally improves auditing?