acceptodds
Under review as a conference paper at ICLR 2027

A Permutation Audit for Cell-Level Foundation-Model Transferability Claims

Abstract

Transferability scores such as LogME are sometimes read as cell-level warnings: a low score for a (foundation model, task) cell is taken to predict that this particular cell will go wrong. The usual evidence is a correlation between the score and cell-level error across a grid of models and tasks. Such a correlation can be produced without any cell-level information, by shared task difficulty, by shared model strength, or by the noise in the error estimate itself. We give an audit for such claims: gates that check that the score is defined on the cells and that the outcome has cell-level variance not dominated by seed noise, then three within-task permutation nulls, each aimed at one artefact, all of which must reject in the claimed direction. The noise null adjusts for each cell's measured seed variation instead of dividing by it; we show that a natural ratio standardisation rejects genuine signal when noise grows with difficulty (pass rate ), whereas the adjusted null does not (1.00). On a synthetic grid the audit passes 0.6% of no-signal runs and none of five leakage regimes, and detects genuine within-task correlations of 0.4 in every run. Applied to LogME, the H-score and N-LEEP on seven foundation-model pools, 11 of 84 warning-oriented audits pass, 10 of them on two medical-imaging pools where two of the three scores pass on the same outcomes; the audit also identifies where it cannot decide: outcomes dominated by seed noise, targets without seed replicates, and grids too small to have power. The implementation returns a structured record that can accompany a new claim.

Primary area:
interpretability and explainable AI
Keywords:
transferability estimation, foundation models, permutation tests, evaluation methodology, confounding, measurement noise, model selection, LogME, medical imaging

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Li2nOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an auditing procedure—comprising three thresholds and three permutation tests—to determine whether transfer scores contain information specific to task pairs that cannot be explained solely by task effects, model effects, or noise from random seeds. The study evaluates various confounding factors and failure scenarios through synthetic experiments. Within a collection of seven base models, LogME, H-score, and N-LEEP passed the audit 11 times out of 84 trials.

Critical strengths

The paper distinguishes between general correlation and task-specific information, examining the respective influences of tasks, models, and noise. It provides actionable auditing procedures, thereby helping to raise the evidentiary standards for claims regarding transfer scores.

Critical weaknesses

The default audit process provides inadequate control over error rates. As shown in Table 5, the error pass rate ranges from 6.2% to 8.4% under conditions of heteroscedasticity—a phenomenon also significantly present in real-world medical datasets. Furthermore, the conditions for applicability require further clarification, particularly regarding dependencies within model families, the handling of missing data, and the thresholds used to manage noise.

Questions for authors
  1. Why is the condition ($ N \leq D $) sufficient to rule out all three categories of scoring methods?
  2. Do the findings obtained from the 11 successful audits still hold when using held-out models or tasks, conducting independent replications, or employing alternative ways of constructing outcome variables?
Reviewer_myyAOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an audit for the claim that a transferability score such as LogME works as a cell-level warning, meaning that a low score for one foundation model and task pair predicts a large error for that pair. After gates on estimability, interaction variance, and seed noise, three within-task permutation nulls named WITHIN, PROJ, and NOISE target task difficulty, model strength, and the noise channel, and a claim passes only if all three reject. The audit is validated on a synthetic harness and applied to LogME, the H-score, and N-LEEP on seven foundation model pools, where 11 of 84 warning-oriented audits pass, 10 of them on two medical imaging pools.

Critical strengths
  1. Table 2 gives a useful warning for practitioners by showing that dividing the error by its seed noise floor removes genuine signal when noise grows with difficulty.
  2. The supplementary material includes the audit library, the harness code, and the records of all real-data audits, with a README that maps each table to its code.
Critical weaknesses
  1. The paper defines a cell-level claim as a claim about the interaction γ\gamma and treats a score that tracks model strength α\alpha as model-family confounding, which PROJ rejects. Yet the paper itself notes that transferability scores are evaluated by how well they rank models on one task, which largely rewards the model-strength differences that PROJ removes. The audit therefore fails scores for the property they are designed to have, so a PROJ failure such as on the audio pool does not show that a score is useless as a warning.
  2. The overall writing of the paper is poor. In the introduction, two short paragraphs of motivation are followed by a contribution list that packs in the method and uses terms such as gates, within-task permutation nulls, conjunction in a pre-declared direction, and warning-oriented audits before any of them is defined. The list also mixes levels, since the estimability gate in the third item is one of the gates of the audit in the first item, and the fourth item reports results rather than a contribution. As a result, a reader reaches Section 3 without knowing what problem the audit solves or why it matters, so the introduction and the contribution list need to be rewritten.
  3. The audit outputs a pass or fail label, but no experiment checks whether this label is correct or useful. The synthetic experiments test it only on artefacts that the nulls were built to catch, and the real-data experiments report labels without any ground truth, such as whether scores that pass warn better than scores that fail on held-out models or tasks.
Questions for authors
  1. Section 5 reports that without the estimability gate only 3 warning-oriented audits pass and nine medical passes turn into failures, so the main positive result depends on removing 11% and 28% of the cells on the two medical pools. Appendix C computes N-LEEP on PCA-reduced embeddings, and a shrinkage estimator of the H-score covariance has been proposed for the same high-dimensional problem [1]. Do the medical passes survive when LogME and the H-score are computed with PCA or shrinkage instead of dropping cells? If they do, the gate is not needed for the result, and if they do not, the positive conclusion rests on which cells are removed.

[1] Newer is not always better: Rethinking transferability metrics, their peculiarities, stability and performance. ECML PKDD, 2022.

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

This paper proposes an audit for claims that transferability scores identify model–task pairs with large downstream prediction errors. It removes task and model effects, adjusts for estimated seed noise, and combines three permutation tests with gates for estimability, interaction variance, and noise dominance. Synthetic experiments test several confounding scenarios, and an application to three scores across seven foundation-model pools finds 11 passing warning-oriented audits.

Critical strengths

I like the distinction between predicting task difficulty or model strength and identifying something specific to a model–task pair. The synthetic experiments make these different sources of association concrete. This is a useful evaluation question, and the structured verdicts help distinguish insufficient evidence from a failed test.

The noise-adjustment analysis is interesting. The comparison with dividing errors by their estimated noise level shows how an apparently reasonable normalization can remove the signal being tested. I also appreciate the reporting of low-power settings, noise-dominated outcomes, and cases where the audit’s calibration deteriorates.

Critical weaknesses

The estimability gate needs a stronger justification, and it materially changes the findings. Section 3.1 treats (N \leq D) as implying that LogME is undefined. But feature dimension alone is not sufficient for that conclusion. The LogME authors’ extended analysis proves that padding features with zeros leaves the score unchanged, while the proposed gate can change from retaining to excluding the same cell after such padding. This gives a concrete counterexample to the gate as stated. The covariance issue for an unregularized H-score also does not establish the same restriction for LogME or PCA-based N-LEEP. This matters because §5 reports only three passing audits without the gate, with nine medical passes becoming failures. I’d want score-specific validity checks, a distinction between numerical failure and mathematical non-estimability, and the main results rerun under those checks. See You et al., Ranking and Tuning Pre-trained Models, Corollary 4. JMLR 2022

I’m not yet convinced the audit provides the false-pass control needed for its intended use. The intersection–union argument is valid conditional on valid component tests, but it does not establish that the proposed permutations and gates satisfy that condition. Table 5 reports false-pass rates up to 8.4% under heteroscedasticity at a nominal 5% level. Section 5 then reports 3–7-fold variation in model-specific error spread on the medical pools, correlated with score variability. The fact that all passes survive per-model standardization is useful, but I’d like to see null calibration of that modified procedure itself. The noise-dominance gate is also calibrated on a particular synthetic family; Appendix C acknowledges that it can miss noise-dominated interactions when model and task effects inflate the median outcome. For a paper proposing an audit, these are central validity questions.

Questions for authors

Does (D) mean ambient embedding dimension or effective rank? How does the estimability gate handle zero-padding invariance, and how many conclusions survive score-specific checks that retain valid LogME and N-LEEP cells?

Can you report false-pass rates for the per-model-standardized procedure under heteroscedasticity, missing cells, and correlated model families? Do the medical findings survive a permutation scheme calibrated for these dependencies?

Can you clarify how the learning-curve baseline is fitted and evaluated, including whether each audited cell’s outcomes participate in fitting its prediction? This would help establish what its residuals and estimated noise floors represent.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.