acceptodds
Withdrawn from ICLR 2027

From Density to Detection: Diagnosing Likelihood Failures and Recovering OOD Signals

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Likelihood-based out-of-distribution (OOD) detection assumes that high model density provides evidence of domain membership. Simple OOD images challenge this assumption across flow and autoregressive models. We examine the resulting failures at the level of individual-image decisions. Within-ID rankings reveal strong likelihood–compressibility alignment in Glow, PixelCNN++, and iGPT. Equal-entropy controls separate whole-image distribution entropy from visual complexity; training from scratch on complexity-filtered images then shows that excluded images outrank allowed images in of Glow comparisons. Noise-only training further reveals structured high-score preferences under a known population density. These controls distinguish limitations of an oracle-density rule from errors in the density learned by a model. We develop the consequences for detection through an –– decomposition: representation density, local volume change, and conditional variation along representation fibers. This framework extends the flow decomposition to general noninvertible networks and motivates learning from their component responses. Across 40 classifiers, the resulting LJF detector improves AUROC over its LF ablation by , , and percentage points on CIFAR-10, CIFAR-100, and ImageNet. Adding geometry also improves a strong 27-score fusion in 212 of 229 model–task comparisons under source-matched calibration. The findings clarify why raw density can mislead OOD decisions and show how its underlying model responses retain useful detection information.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Out-of-Distribution Detection, Classifier Geometry, Jacobian Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_aFm5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper addresses likelihood-based OOD detection. Controlled diagnostics on generative models (Glow, PixelCNN++, iGPT) show that likelihood favors simple images, which the authors attribute to model misestimation rather than to entropy or typical-set effects. The paper then proposes an L–J–F decomposition of the log-density (representation density, Jacobian volume, fiber term), instantiated on frozen classifiers and combined by a logistic regression trained on labeled ID and OOD data. Experiments cover 40 classifiers on CIFAR-10, CIFAR-100 and ImageNet, compared with 27 existing OOD scores and their fusion, plus a flow-based counterpart on CelebA vs. SVHN.

Critical strengths
  • Diagnising & Understanding likelihood failures in OOD detection is a relevant, long-standing question.
  • Some diagnostics are carefully controlled. In particular, the excluded-support experiment cleanly separates model misestimation from typical-set explanations.
  • The classifier evaluation is broad (40 models, six density estimators), with a fusion baseline of 27 score families.
Critical weaknesses
  • Presentation and clarity. The paper is difficult to follow, which is a serious issue for a top-tier venue.

    • The method is never stated plainly in the main text. There is no contribution list and no overview figure. The actual detector, a logistic regression on three scores, only appears explicitly in the appendix (Eq. 30).
    • The text is saturated with defensive qualifications (“this does not establish…”, “this is not a claim that…”), which bury the main message.
    • Notation is overloaded: L and J denote different quantities for flows and for classifiers. Several terms are used without definition (“source-matched calibration”, “readout”, “component footprint”, “higher-score fraction”).
    • Space is poorly allocated: elementary identities (Eqs. 2–3, the Gaussian example) take main-text space, while the estimators and the evaluation protocol are relegated to the appendix.
  • Unclear focus: generative diagnostics vs. a classifier detector. The paper combines two largely independent contributions.

    • The diagnostics concern the simplicity bias of generative likelihoods.
    • The proposed detector operates on classifiers and uses no generative likelihood, so the diagnosed failure plays no role in it.
    • The coarea decomposition is the only bridge, and it is not operational for classifiers (next point).
    • The flow-side detection results rest on a single setting (CelebA vs. SVHN).
  • The decomposition is not what is estimated. The identity log p_X = L + J + F holds for a specified input density p_X, which a classifier does not provide.

    • L and F are Gaussian fits in feature space, whereas J is computed in input space, so the three terms do not refer to the same density.
    • F is fitted unconditionally on a feature residual, while the exact F is conditional on the input fiber. The paper’s own Appendix F.1 (Eq. 20) shows that a feature-space conditional does not identify the input-space F.
    • Consistently, the fixed sum of components does not work (CIFAR-100 AUROC 63.18 for J+L+F vs. 63.56 for L+F); only the learned weighting helps. The decomposition is therefore a motivation rather than the basis of the estimator.
  • Limited novelty.

    • Most diagnostic findings confirm known explanations: the compressibility link (Serrà et al., 2020), the density-vs-typicality distinction (Nalisnick et al., 2019b; Le Lan & Dinh, 2021), and model misestimation (Zhang et al., 2021).
    • On the flow side, the radial recalibration is borrowed (Dheur & Ben Taieb, 2025). The two-sided test is a conformal version of typicality testing. Component-based flow detectors already exist (DoSE, Morningstar et al., 2021; likelihood-free detection, Ahmadian & Lindsten, 2021).
    • On the classifier side, L resembles a Mahalanobis score on logits and F a ViM/Residual score, leaving the Jacobian term J as the only new ingredient.
  • Reliance on labeled OOD data from the test source. The detector’s weights and signs are learned on labeled OOD images from the same source as the test OOD images.

    • This turns OOD detection into binary classification against a known distribution, whereas the defining difficulty of OOD detection is that test-time shifts are unknown: OOD samples are by definition anything outside the ID distribution and cannot be collected representatively. In my view, this is a dead end by construction: a detector tuned to one OOD source offers no guarantee for the next one.
    • Each OOD dataset is split into calibration and test halves, so the detector is trained and evaluated on i.i.d. samples of the same OOD dataset. At the detector level, there is no distribution shift at all.
    • The fixed-sum results above show that the components only help once their weighting is learned on the target OOD, which is precisely the information unavailable at deployment.
    • The strongest baseline (ALL) is fitted in the same regime, so the reported gains hold only in this unrealistic setting. An OOD-free variant of LJF, compared with standard unsupervised detectors, is needed. At a minimum, a cross-source evaluation (e.g., leave-one-source-out) would show how much of the gains depends on source matching.
  • Weak experimental results.

    • Even in this favorable supervised setting, LJF alone remains below the best single baseline selected on calibration data: 88.87 vs. 90.65 AUROC on CIFAR-10, 82.30 vs. 82.94 on CIFAR-100, and 80.57 vs. 88.13 on ImageNet. The LJF-over-LF gains are thus gains over a weak ablation.
    • The only improvement over existing methods is +0.39 AUROC when adding J to the 27-score fusion. It is concentrated on a few easy targets (e.g., SVHN for CIFAR-100) and negligible elsewhere (e.g., +0.01 on NINCO, +0.03 on OpenImage-O for ImageNet).
    • On ImageNet, the LJF-over-LF interval for the primary estimator includes zero.
    • The cost of computing the full input-to-logit Jacobian, which forces an arbitrary 32-dimensional logit projection on ImageNet, is not discussed.
Questions for authors

See weaknesses

Reviewer_1YUqOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper explores the issue of likelihood-based density estimation applied to out of distribution (OOD) detection. The authors first propose a global analysis of high likelihood scores for detection. They show that high likelihood does not always characterize in distribution (ID) data; they study whether entropy is linked to the visual complexity of the image; they check that a high-density region can contain very few points, which can lead to high likelihood scores; they finally inspect the effect of training on natural images. Then, in response to this analysis, they propose to decompose the density of the data through the coarea formula into three terms (described as representation frequency, local transverse volume change and variation among inputs that the representation merges) and compute a statistic motivated by each term; these statistics are combined by a logistic regression trained on ID and OOD data. Eventually, they present experimental results to check the importance of the second term JJ, i.e., the volume change, for OOD detection, by comparing the proposed decomposition with and without JJ and do the same by adding JJ to an aggregation of known OOD detectors.

Critical strengths

The paper addresses an important problem for OOD detection: the high-density region curse and the role of image complexity of training data are known issues.

Critical weaknesses

The paper is poorly written. Several ambiguous notions are used without proper definition, such as "gaussian base" l253, or more importantly the notion of "flow", which is introduced as "For coupling-layer flows, Kirichenko et al." with a reference about normalizing flows, which is never stated in the paper. Similarly, Glow is never described or associated with its paper in the main text. This is, in fact, done in Appendix C, which is never cited in the main, like appendices A, D and E. Other notions are only defined in the appendix, such as p−p_{-} and p+p_{+}: appendices should be cited precisely when employed. Also, Appendix F contains only F.1. The same holds for experiments: Table 3 is never mentioned in the text, although it is placed in the main paper (and probably described in section 5.4, paragraph 2 ?). In this section 5.4, "CelebA face-model experiments" are mentioned, while no table with CelebA results exists (maybe in appendix E, where celeba is mentioned?). Overall, the paper is very difficult to read and understand, while the notions it uses are not that difficult.

The second weakness is the value of the contribution. Even though the idea of explaining classical difficulties in detection is relevant, it has already been addressed by the works mentioned in the paper, and thus what this paper could have brought is clarity on the problem, which, as previously stated, is not the case. The proposed detector aims to study whether the "volume change information" is useful for detection. In the end, the results do not seem striking. Moreover, the estimators are only justified in the appendix, while this is an important point of the discussion.

Questions for authors
  • The estimator of JJ seems central to the discussion: can you justify here the choices made for it (and also for L and F)? Could you check experimentally how close the estimator is to JgJ_g? And can you clarify the role of each "JJ", e.g., JJ, JgJ_g, JεJ_ε, plus the ones in appendix F.1 : JhJ_h and JWJ_W ?

  • Can you justify the choice of the methods included in ALL? Do any of them use the same information as JJ?

Reviewer_LoEgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper examines the resulting failures at the level of individual-image decisions. Within-ID rankings reveal strong likelihood--compressibility alignment in Glow, PixelCNN++, and iGPT. their proposed Equal-entropy controls separate whole-image distribution entropy from visual complexity.

Critical strengths
  1. this paper is well written and easy to follow

  2. the figures are informative

  3. experiments are conducted on various scales of datasets

Critical weaknesses
  1. In Sections 4.3 and 5.1, the authors explicitly claim that their detector is trained using labeled target OOD examples from the same named sources subsequently used for testing, which leads to unfair comparision with post-hoc ood detection methods

  2. Computing input-to-readout Jacobians and their singular spectra are computationally expensive. However, the manuscript lacks efficency analysis.

  3. the compared baselines are out-dated. the authors are expected to compare their method with more recent baselines.

  4. it can be found that the third tem of Eq. (8) calculates the variation among all inputs producing the same output. However, the authors' implemetnation only is based on the variation in the feature space. since the difference between the both can not be addressed by neither the Gaussian consistency check nor Jacobian regularization, it is unclear whether the performance comes from the decomposition of Eq. (8) , therefore deteroriating the contribution of this paper.

  5. the authors are expected to isolate and substantiate its contribution beyond [1,2].

[1] Density of States Estimation for Out-of-Distribution Detection, AISTATS2021

[2] Understanding Failures in Out-of-Distribution Detection with Deep Generative Models, ICML2021

Questions for authors

see weakness

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

Related papers

Discussion (0)

Sign in to comment.