acceptodds
Withdrawn from ICLR 2027

Does High Accuracy Imply Trustworthiness? A Large-Scale Evaluation of Confidence and Risk in Frozen Vision and Video Foundation Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

As large-scale vision and video foundation models enter high-risk settings, a standard safeguard is selective prediction, where a model acts only on high-confidence outputs while deferring uncertain ones. However, accuracy alone does not reveal whether a model’s most confident predictions are reliably correct. Common uncertainty metrics also do not directly address this issue. AUROC and expected calibration error (ECE) evaluate confidence across the full prediction set, FPR@95 evaluates at a fixed error-rejection target, and other metrics remain confounded by model accuracy. In this work, we focus on the trustworthiness of a model's confidence levels among predictions selected for action. We propose that the positive likelihood ratio (LR+), originally defined for diagnostic testing, can be adapted to measure how much more often correct than incorrect predictions are included among a fixed fraction of the most-confident predictions. We evaluate LR+ across 18 frozen vision and video foundation-model backbones from six families and three scales, spanning image classification, semantic segmentation, video classification, and action anticipation. On ImageNet-1K, DINOv3 ViT-H+ and SigLIP2 g-opt achieve nearly identical top-1 accuracy (88.4% vs. 88.1%), yet their LR+ values among the top 20% most-confident predictions differ by more than twofold (24.0 vs. 49.7). Across all 18 models, AUROC and FPR@95 show little variation and rank models differently from LR+, which reveals substantial differences. Across downstream tasks, the relationship between accuracy and LR+ changes, and different models achieve the highest LR+. These results show that the trustworthiness of a model based on its confidence scores cannot be inferred from accuracy alone or assumed to transfer across benchmarks, and should be evaluated for each task at the fraction of predictions intended for action.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Uncertainty, Vision and Video Foundation Models, Trustworthiness

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_8sHtOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether high predictive accuracy implies trustworthy confidence in selective prediction systems. The authors introduce the positive likelihood ratio (LR+), adapted from diagnostic testing, to measure the relative retention of correct versus incorrect predictions at a fixed coverage level, arguing that conventional uncertainty metrics do not adequately capture reliability in the high-confidence region. The study evaluates 18 frozen vision and video foundation models from six families across five datasets, spanning image classification, video classification, action anticipation, and semantic segmentation. LR+ is compared against AUROC, FPR@95, ECE, and signed ECE. The main findings are: (1)on ImageNet-1K, DINOv3 ViT-H+ and SigLIP2 g-opt are accuracy-matched (88.4 vs 88.1) but differ ~2× in LR+ at 20% retention (24.0 vs 49.7); (2) AUROC and FPR@95 vary little across the 18 models and rank them differently from LR+; (3) the accuracy–LR+ relationship and the top-LR+ model change across tasks; (4) MC-dropout improves ECE but leaves median LR+ essentially unchanged. Overall, the paper argues that predictive accuracy alone is insufficient for assessing confidence reliability and advocates evaluating confidence at the operating point relevant to selective prediction.

Critical strengths
  1. Relevant and well-motivated research question. The deployment question is well motivated and clearly stated: in an abstaining system, the only predictions that matter are those above the action threshold, and a metric evaluated at that operating point is a reasonable and sensible goal.
  2. Broad empirical evaluation across modern foundation models. The empirical scope is broad for an uncertainty study of frozen backbones: six families spanning three pretraining paradigms, three scales, five datasets, four task types, with a uniform probe protocol and a shared hyperparameter search per task.
  3. Emphasis on operating-point-specific evaluation. The use of LR+ at fixed retention fractions provides an interpretable perspective on how confidence separates correct and incorrect predictions within the selected set. Unlike metrics aggregated across all thresholds, this analysis focuses on an operating region relevant to abstention. The retention curves further demonstrate that model comparisons may change depending on the selected coverage.
  4. Writing is clear. Limitations are stated honestly (single seed, probe-size confound, Diving48 size, label-noise sensitivity).
Critical weaknesses
  1. LR+ at fixed coverage is a reparameterization of selective risk, not an independent quantity. The main methodological contribution is the adoption of LR+ at fixed coverage. However, LR+ can be expressed directly in terms of overall accuracy (a) and selective risk (r(q)): LR+(q) = [(1 − r(q)) / r(q)] · [(1 − a) / a]. Thus, at fixed coverage, LR+ provides no independent information beyond these two quantities. Importantly, a model may achieve lower selective risk yet have a lower LR+ due to differences in overall accuracy. While the authors discuss AURC, E-AURC, NAURC, and SAC (Sec. 2), they do not directly compare LR+ with selective risk at matched coverage or empirically evaluate its added value over existing selective prediction metrics. Consequently, it remains unclear what additional actionable insight LR+ provides, particularly for deployment decisions where the absolute error rate among accepted predictions is critical.

  2. Insufficient evidence that conventional uncertainty metrics conceal meaningful differences. Comparing the range of AUROC (0.864–0.885) with that of LR+ (24.0–53.1) is not meaningful across metrics with different scales; a bounded metric near its ceiling is necessarily compressed, and a ratio is necessarily expanded. Table 6 shows that AUROC, FPR@95, and ECE all favor SigLIP2 with CIs excluding zero, so the conventional metrics agree in direction. The rank-disagreement example (DINOv3 ViT-L: AUROC 0.879 "among the highest", LR+ 24.0 "last") is given without AUROC CIs; with all 18 models spanning only 0.021 in AUROC, it is unclear how many of these ranks are statistically distinguishable. While LR+ may emphasize differences at a specific operating point, its additional discriminative value is not sufficiently demonstrated. A quantitative analysis of model-ranking disagreements, supported by paired statistical comparisons, would strengthen this claim.

  3. Limited evaluation under deployment-relevant conditions. As acknowledged in Section 5, all results are on in-distribution data. Given the safety-critical motivation, evaluating the stability of the confidence threshold and of selective risk under at least one representative distribution shift would strengthen the practical claims.

  4. Limited robustness analysis of probe training. Each probe configuration is trained once; bootstrap intervals cover evaluation-set variability only, not probe initialization or training stochasticity (Sections 3.5, D.3). Running 3–5 seeds for at least DINOv3 ViT-H+ and SigLIP2 g-opt on ImageNet-1K would be inexpensive and would establish whether the central finding is robust.

  5. Figure 4 is unreadable at print size.

Questions for authors
  1. Added value of LR+. Since LR+ at fixed coverage can be expressed directly in terms of overall accuracy and selective risk, what additional information or practical benefit does it provide? Could the authors include direct comparisons with selective risk at matched coverage and with established metrics such as E-AURC, NAURC, or SAC?
  2. Agreement with conventional metrics. Table 6 shows that AUROC, FPR@95, and ECE all favor SigLIP2 g-opt over DINOv3 ViT-H+, consistent with LR+. How often does LR+ change model rankings relative to the conventional metrics, and are those disagreements supported by paired confidence intervals?
  3. Robustness to probe training. How sensitive are the reported LR+ values and rankings to probe initialization and training randomness? Could the authors provide multi-seed results for representative models, in particular the DINOv3 ViT-H+ vs. SigLIP2 g-opt comparison and the within-family scaling trends?
  4. Deployment-relevant thresholds. Thresholds are currently set from evaluation-set quantiles (Algorithm 1). How would the conclusions change if thresholds were selected on a separate calibration set and held fixed during evaluation? Do LR+ and the associated selective risk remain stable under distribution shift?
1 comment

Thank you for the detailed response and additional experimental results. The robustness analyses and metric comparisons answered most of my questions. However, the practical benefits of LR+ over current metrics remain limited. Given the improvements, I will raise my score accordingly.

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

This work identifies the mismatch between high predictive accuracy and high trustworthiness in deep models. It proposes the positive likelihood ratio (LR+) as a measure of trustworthiness and conducts an extensive evaluation across various vision foundation models and tasks. The experiments reveal several notable patterns concerning the relationship between accuracy and trustworthy confidence.

Critical strengths
  • The motivation is clear. Although the mismatch between accuracy and trustworthiness has been studied for many years, an extensive evaluation of vision foundation models remains interesting and potentially valuable.
  • The core idea is straightforward and easy to grasp. Simple yet effective ideas are always welcome in the vision community.
Critical weaknesses
  • LR+ behaves quite differently from other trustworthiness metrics. Although these metrics all claim to measure trustworthiness, this work provides insufficient empirical evidence or analysis to demonstrate why LR+ is a better or more appropriate metric. The paper reports the observed differences but does not adequately explain them.
  • The mismatch between high accuracy and high trustworthiness has been studied extensively, so merely reiterating this issue does not constitute a sufficient core contribution. The paper reports several additional observations, such as “the relationship between accuracy and trustworthy confidence is task-dependent” and “scaling improves accuracy more consistently than trustworthy confidence.” However, it largely stops at describing these empirical phenomena. For example, the paper does not investigate why different model families exhibit different behaviors, what factors underlie these patterns, whether any general principles can be identified, or how trustworthiness might be improved. This lack of deeper analysis substantially weakens the paper’s contribution.

Overall, the paper presents a large collection of experiments but remains largely descriptive. While I learned several empirical facts from the paper, it is unclear what broader insights or actionable conclusions can be drawn from them.

Questions for authors
  • The equations are not numbered.
  • The discussion of “Failure Detection and Calibration” in Section 3.1 mainly covers standard background knowledge and largely repeats the second paragraph of the Introduction. I suggest using the limited space in the main paper for more substantive content, particularly a deeper analysis of the experimental findings.
  • Figure 2 contains too much information, and its small font significantly reduces readability.
  • The notation should be made consistent: the main text uses “LR+,” whereas the equations use “LR” with “+” as a superscript.
Reviewer_cErrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors address the safety of vision and video foundation models. In particular, they emphasize that prediction accuracy alone is not sufficient, information regarding uncertainty and confidence should also be presented. The authors propose that the positive likelihood ratio (LR+), which was originally defined for diagnostic tests, can be adapted to measure how much more frequently correct predictions are found among false predictions in a fixed proportion of the predictions with the highest certainty. They evaluate their LR+ method using 18 backbones for image and video base models from six families and three scales covering the areas of image classification, semantic segmentation, video classification, and action prediction. The authors compare their metric with other well-known metrics in the field of uncertainty quantification and failure detection.

Critical strengths
  1. The topic addressed by the authors is highly relevant, as computer vision tasks are frequently used in safety-critical fields such as medicine and autonomous driving. The issue of trustworthiness has not yet been systematically evaluated for the current generation of large-scale vision and video foundation backbones. In addition to well-known metrics such as ECE and AUROC, the authors also introduce LR+.

  2. The authors test six different state-of-the-art backbones in various configurations, using five different datasets. They also investigate the network's direct output (MSP) and analyze MC-dropout uncertainty. Overall, it is a good first step toward a comprehensive analysis of the trustworthiness of vision foundation models.

Critical weaknesses
  1. I find some of the statements made by the authors to be a bit too vague. "Therefore, these metrics either remain influenced by overall accuracy or do not directly evaluate trustworthiness at a chosen fraction of the most-confident predictions." This statement is a bit too broad, since ECE and AUROC measure different properties and are not both equally influenced by overall accuracy. Furthermore, just because a metric does not explicitly focus on a specific proportion of high-confidence predictions does not mean that it cannot, in principle, assess their trustworthiness. "LR+ directly addresses both limitations described above." This claim seems overstated, since LR+ does not directly quantify the absolute error rate for retained predictions and, when evaluated using a single retention rate, does not characterize reliability across different coverage levels. While it may complement existing metrics, it does not necessarily address both of these limitations on its own.

  2. I am missing a few points on the experiments. Only one dataset was tested per task. LR+ looks at the top 20%, while ECE and AUROC cover the entire range of predictions; perhaps it would make sense, for example, to compare it with ECE only for the top 20%. I would like to see more discussion of the experiments, for example, are there differences between the three model scales, or regarding the various datasets/tasks?

Questions for authors
  • Could you explain the points mentioned in “Weakness 1” in more detail?
  • "Point comparisons require at least 20 observations in each group" Why exactly 20? Is there a theoretical basis for this, or was it determined empirically?
  • These absolute spans (comparing AUROC, LR+ ...) are not directly comparable due to the different scales. Would relative values be more fair?

Remarks:

  • I would not name chapter 3 “Methods,” since that sounds as if a new method is being introduced. It is more about the experimental setting.
  • At the beginning of Chapter 3, MSP is introduced as a measure of uncertainty, but MC-Dropout is only introduced in Section 3.4, it's a bit confusing.
  • The labels on the figures are very small.

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

Related papers

Discussion (0)

Sign in to comment.