acceptodds
Under review as a conference paper at ICLR 2027

Certified Out-of-Distribution Safety for Generative Maps from In-Support Geometry Alone

Abstract

Every generative model is trained on a finite set of examples, and every generative model is eventually asked for something a little outside that set. What happens next is usually found out the hard way: you sample, you look, and you decide whether the output is still trustworthy. The awkward part is that checking whether a model behaves outside its training support seems to demand data from outside that support, which is exactly the data nobody has. We show the check can be done entirely from the inside. Near the edge of a model's support we fit a small atlas of local linear approximations to the map and read off three quantities that never require leaving the support: how well the atlas already fits, how rough the map looks at the finest scale the atlas can see, and how sharply it bends. Those three numbers name a radius: a distance beyond the edge within which the map's extrapolation error is guaranteed to stay under a tolerance the user picks. We test the prediction against ground truth on synthetic warped manifolds, on maps fitted to real tabular data, and on flow-matching networks trained on three image datasets. On the synthetic families the predicted radius sits inside the true one on every cell we measured. The certificate is conservative rather than optimistic, and tight enough to be useful rather than vacuous. On trained networks it does too, but only after one constant is recalibrated for the model class, and we report exactly what goes wrong when it is not. The algebra underneath is machine-checked in a proof assistant, so the parts that are theorems are theorems. Two findings surprised us. The bending term must be measured as the curvature of the map itself and not the curvature of the surface it draws. The second, more natural choice quietly smuggles in a dependence on how much the map contracts, and breaks the guarantee. And the constant tying the two roughness measurements together, genuinely constant on synthetic maps, becomes a per-model quantity with a heavy tail on trained networks, so it must be calibrated per model class and at a high quantile rather than at a typical value. We also report where the method stops: it certifies smooth deterministic samplers, and a diffusion-transformer sampler we tried falls outside that class for reasons we measure rather than guess.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Out of Distribution, OOD, Extrapolation, Deep Learning, Theory

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_bJ6XOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Generative models are trained on data drawn from a given distribution, yet they are routinely queried outside their training support, where their behaviour is unspecified. This paper proposes a certificate predicting how far outside its support a generative model remains close to its own linear extrapolation; according to the authors, beyond that point the model departs from its in-support behaviour and becomes unpredictable.

In a band near the support boundary, the map is approximated by an atlas of local affine charts. From this atlas, three quantities measured inside the support (the atlas error at the boundary, the affine residual and the curvature), together with a calibrated constant κ\kappa, define an error model quadratic in the distance, whose positive root gives a radius dsafed_{safe} below which the deviation between the map and the atlas should remain under a tolerance τ\tau. The algebra of the formula is verified in Lean, while the validity of the radius rests on the empirically tested assumption that the error model dominates the true error. The radius is safe on a synthetic testbed. On 2D tabular maps and flow-matching models with a 2D latent, however, it overestimates the true radius on part of the cells, and κ\kappa must be recalibrated for each model class.

Critical strengths
  • Transparent and disciplined experimental protocol. Discovery and fresh cells are kept separate, constants are fixed before validation, and failures are reported rather than tuned away. This is exemplary and it is also what makes a precise assessment possible, since the method's weaknesses can be read directly from the authors’ own tables. However, it does not compensate for the conceptual issues below.

  • A clean falsifiable test of the curvature estimator (section 4.2). The prediction that the certified radius should be invariant to the map’s affine part is tested in a controlled setting and clearly discriminates the raw map-Hessian from the normal-projected second fundamental form. This is the best-supported finding of the paper, although its scope is limited to a single synthetic family.

Critical weaknesses
  • The certificate offers no guarantee, and the quantity it predicts can be computed directly. Soundness (Proposition 2) holds only under the assumption E(d)≤E^(d)E(d) \leq \hat{E}(d), which is neither proved nor provable without constraints on F outside the support. The authors acknowledge this in Appendix A, yet the title, abstract and Lean formalisation suggest a guarantee the paper does not provide. dsafed_{safe} is therefore an empirical predictor, and must be judged as such against the obvious alternative: the bounded quantity E(d)=∣∣Fatlas−F∣∣E(d) = || F_{atlas} - F || needs no ground-truth data, and the paper itself obtains dmeasd_{meas} by simply querying F on outer shells, at a cost comparable to fitting the atlas. However, the bounded quantity E(d)E(d) needs no ground-truth data: the paper itself obtains dmeasd_{meas} by simply querying F on outer shells, at a cost comparable to fitting the atlas. Moreover, the predictor is less reliable than this direct estimate κ\kappa must be calibrated on out-of-support error curves, and with the transferred κ\kappa the certificate overclaims on all three image datasets, with violations remaining on Fashion-MNIST and EMNIST even after in-sample per-class calibration. I cannot see what dsafed_{safe} offers over direct estimation. A comparison in cost and reliability would be required, and the current evidence suggests dsafed_{safe} would lose on both. This is the main reason for my recommendation.

  • The bounded quantity measures regularity, not reliability. The certificate bounds the deviation of F from its own linear extrapolation, not the quality of its outputs. For exemple, a generator that stays on a curved data manifold must deviate from its tangent extrapolation, so a large E may reflect good behaviour. The paper provides no evidence linking E(d)E(d) to output quality (sample plausibility, distance to the data manifold, out-of-range pixels). The framing as "OOD safety" is therefore not supported, even if the predictor itself were accurate.

Questions for authors
  • Since dmeasd_{meas} is obtained by querying F on outer shells, with no ground-truth data, is there a practical setting where F can be queried inside the support but not outside, or where outer queries are substantially more costly? Please report the number of F evaluations required to compute dsafed_{safe} (including atlas fitting and κ\kappa calibration) versus direct estimation of dmeasd_{meas} to comparable precision.

  • Calibration of κ\kappa without OOD queries. How would a practitioner set κ\kappa for a new model class without computing out-of-support error curves? If such curves are required, in what sense is the certificate obtained "from in-support geometry alone" ?

  • Do you have evidence that E(d)E(d) correlates with any measure of output quality as d grows, for example sample plausibility, distance to the data manifold, or fraction of out-of-range pixels?

  • Precision of dmeasd_{meas}. On Fashion-MNIST and EMNIST, E(d)E(d) is a 99th percentile estimated from 60 samples per shell, effectively a maximum. Can you report confidence intervals on dmeasd_{meas}, and indicate whether the observed violations (or their absence) are robust to this estimation noise?

Suggestions for future work

  • Determine conditions under which E(d)≤E^(d)E(d) \leq \hat{E}(d) holds : this would turn the predictor into a genuine conditional certificate.

  • A systematic study of the relation between E(d)E(d) (or dsafed_{safe}) and the quality of generated outputs as d grows, e.g. sample plausibility, distance to the data manifold, or fraction of out-of-range pixels. This would establish whether the certified quantity is a meaningful proxy for OOD reliability.

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

The authors define an atlas on the ``edge'' of the manifold from which training data for a generative model come from. They define some heuristics based on the smoothness of the atlas that determine when the generated samples can be trusted as coming from the true data distribution.

Critical strengths

OOD detection just from training data is an important open problem with wide applicability wherever generative models are used.

Critical weaknesses
  1. Starting with the abstract, there are many sentences that are not mathematically meaningful and therefore obstruct any intended meaning of the paper. In other words, the paper's results are unintelligible. For example, " it certifies smooth deterministic samplers, and a diffusion-transformer sampler we tried falls outside that class for reasons we measure rather than guess."

  2. Many differential geometric terms are used in a non-mathematical way: "We show the check can be done entirely from the inside. Near the edge of a model’s support we fit a small atlas of local linear approximations to the map and read off three quantities that never require leaving the support:" what exactly does an edge, radius, bending etc mean?

  3. The radius is defined using fitted second derivatives. Is this is a computable or an analytical quantity. Either way, it is hard to reconcile how this helps one with OOD detection using only the training data "on the inside"

Questions for authors

Please see the weaknesses above. There are many works that deal with the geometry of the data manifold (e.g., Pidstrigach 2022, Chandramoorthy and de Clercq 2025) and support robustness of diffusion models in practice. Your results seem to suggest that, because of support robustness and because in practice, we do not know the data manifold exactly (i.e., we do not have a parameterization), we can take samples near the data manifold to be approximately from the data distribution. This is already implied also from convergence in TV/Wasserstein distance of diffusion models (e.g. Lee et al 2023). What new results are sought in the present paper? How are they useful in practice for OOD detection from training samples?

Reviewer_2kcdOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a new method to set a specific radius (in the noise space) so that sampling noise inside the radius would suppress the out-of-distribution samples up to the user’s tolerance level. First, the paper derives a formula to estimate extrapolation error. The method first samples charts (in the noise space) around the boundary shell of the training data’s support, then computes the affine and quadratic mappings that approximate the local mapping of the charts. The constants in the formula come from measuring the error and residuals of those mappings. The safe radius is the root of the equation that equates the formula and the user’s tolerance. The paper presents some machine-verified statements regarding the method's properties. The experiments justify the design choices, empirically verify some properties, and measure a few real datasets.

Critical strengths
  1. The paper attempts to develop a method on a theoretical foundation based on geometric measurements. If the paper can demonstrate the method’s practical value, the theoretical foundation would also help in understanding the nature of OOD samples.
Critical weaknesses
  1. The paper is generally difficult to follow even with careful reading. The motivation and practical value are unclear. The figures include too many redundant details, such as dotted lines and text, forcing readers to repeatedly switch between the figures and the paper’s description. The authors do not clearly describe the purpose of the proofs and experiments, making it hard to understand what they intend to emphasize.
  2. The experiments focus on showing the required properties of the proposed idea rather than demonstrating the method's practical value. Some experiments only justify why the proposed method uses a specific component/hyperparameter. Those results do not demonstrate any practical value of the proposed idea, so we can move them to the Appendix. The real-data experiments are very limited and provide only numbers that readers cannot judge as good or bad. The paper also raises practical concerns about these limitations, but it does not discuss how to address them or provide additional experiments.
Questions for authors
  1. Please improve the paper's overall presentation.
    • The motivation is unclear. Section 1 states that practitioners would need to query out-of-distribution samples and measure quality. However, it is unclear why this is such an important problem and what benefit the proposed method provides. If this procedure is costly and the proposed method reduces that cost, how much does it cost, and how much can we save?
    • Section 1 lists six contributions, but I don’t think many contributions in one paper are generally a good thing. It only looks like the authors can't prioritize the findings and just want to say “we did many things.”, and presenting all findings in a given page limit usually results in jumping around random topics, making readers more confused. Please be selective about which results to emphasize, and clearly describe why the results are important. This could be about the impact of the theoretical results or the practical implications of the proofs and experiments.
    • At the beginning of Section 3.4 and Section 4, state the purpose of the section. The purpose statement should answer several questions: Why is proving those statements essential, what are the key questions that each experiment tries to answer, why are those key questions important in practice, and eventually what are the key takeaways (for the readers) from the proved statements and the experimental results?
    • For real-data experiments, please include all the details of the models used to fit the generative maps.
    • Simplify the figure and write more thorough descriptions of what the figure means.
  2. This paper proposes a new method to help practitioners, so it is important to demonstrate the method’s practical benefits from experiments. Regarding the paper’s practical value,
    • Include more real-world datasets that would require larger source dimensions.
    • While Figure 17 shows the method's limitations (by providing cases where it does not apply), it also helps readers understand the paper's position and identify who should be interested in using this method in practice. Moving Figure 17 to the paper’s main body would be unnecessary, but the paper should clearly emphasize the scope at the beginning.
    • The wide dispersion of κ\kappa is definitely a practical concern. This value shows large variance even for simple MNIST variant datasets. This might require localized per-cell certificates rather than a global certificate computed from aggregated values. I also feel that the conservativeness of the certified radius comes from aggregation, which may overly avoid sampling where the manifold's local structures are hard to fit.
    • Some subsections contain information not important to practitioners. Section 4.1 mainly justifies the linear-floor model by comparing it with the pure-curvature form. Section 4.4 discusses only why a high quantile should be used for aggregation. Such an exploration of design choices is generally not a practitioner’s concern, so it can be moved to the appendix.
    • One main problem is that the provided numbers cannot be compared to competitors. However, it could be possible to test the reliability of the certified radius in practice. For example, within a specific tolerance level, samples from sources in the certified radius should not be OOD, so state-of-the-art OOD detectors should not flag them as OOD. Also, intentionally measuring the radius on some ill-fitted generative models (prone to generating OOD samples) would provide another comparison and show that the method can compute the certified radius even against ill-fitted models.
    • Propose a real-world scenario that emphasizes the method's practical benefit. Describe the scenario in the introduction. If an experiment can demonstrate the current problem, it would also help convince readers.
    • The experimental results section mentions diffusion transformer as one of the map family. This looks like the paper ran experiments on diffusion transformers, but it only refers to the limitations section and says such a model is out of scope. If they are out of the experimental coverage, don’t include this in the paragraph.
1 comment

Thanks to the authors for their effort on the rebuttal.

I briefly reviewed the response and the revised version, and here are some thoughts.

  • I disagree that the revised version improved the overall presentation and readability.
  • I asked to be more selective when choosing the main contribution. Merging those categories into fewer categories keeps all of them in this paper, so I still don't see the "important" finding the authors want to emphasize. The first sentences of the contributions are just another list of jargon specific to this paper, so it is hard to understand what those contributions are.
  • The motivation is still very unclear, and there are too many missing details about the scenario (in Section 1) that is claimed to be "concrete".
  • I don't think that the new figures are easier to interpret. The new figure introduced more "redundant" text that I pointed out in the initial review, with nearly no margin between lines and text inside the figure. If conveying lots of information (a figure that is full of text and many plots) is necessary, the authors should allocate more space for those figures.
  • The main body of the paper still does not present the necessary parts, and it only refers to the Appendix whenever details are needed. Unless the paper provides enough motivation, general readers will not go to the Appendix to check details. (As far as I know, even reviewers are generally not mandated to do so.)
  • I cannot understand the rationale for adding more details in Section 3.4 because this is the section that readers would not be interested in. The proven statements are specific to the proposed methods, so they do not provide generalizable theoretical findings. Some findings would help justify a design choice, but for that purpose, listing informal findings would be sufficient.

Now I'm fairly confident in my evaluation of score 1. In my best judgment, the paper's quality is well below the acceptance threshold.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.