acceptodds
Under review as a conference paper at ICLR 2027

A Density-Ratio Readout Theory of Label-Free Out-of-Distribution Detection

Abstract

Out-of-distribution (OOD) detection is not simply a test of nominal atypicality. Reference typicality need not reveal which distribution a sample was drawn from. An unlabeled test batch adds population-level evidence of how the test distribution differs from the reference. This paper asks what becomes detectable from this information and how much can be recovered from finite data. At the population level, the batch-to-reference density ratio identifies where the test population is overrepresented relative to the reference, providing direct evidence of distributional difference. Finite data, however, limit how finely this ratio can be resolved, and finer resolution requires higher-order batch moments. Low-degree readouts that maximize the deflection, a standardized batch–reference contrast, have closed forms in the batch moments under a whitened Gaussian working model, with degree 2 requiring only the batch mean and covariance. Experiments from CIFAR to ImageNet show that the simple degree-2 readout outperforms ratio estimators on average, with its near-OOD gain largely reproduced by the directions along which the batch covariance departs from the reference.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
out-of-distribution detection, density-ratio estimation, batch-based OOD detection

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_TFwHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The work proposes a detection method based on a density ratio, an approach motivated by the fact that if anomalies fall in a high density region of the reference distribution, a detector using only the reference distribution may treat them as in-distribution. The proposed detector compares a reference dataset, with in-distribution data, to a batch dataset, made of both in-distribution and anomaly data; its resolution is described by the polynomial degree of the score, being the expressive moments of the batch dataset. A closed form expression of the scores for degrees 1 and 2 is derived for a whitened Gaussian as reference distribution. The method is then tested on classical datasets (cifar10, svhn,...), where the results tend to favor the proposed "ratio" detector of degree 2 based on the mean and covariance of the batch vs in-distribution data, in particular on near out of distribution data.

Critical strengths

The proposed methods is a well constructed aggregation of the state of the art, with a clear segmentation of the different contributions: the link from the batch dataset literature (Blanchard et al...) and the polynomial readout is a good idea.

The derivations seem correct, the proposed method performs well, and its results are properly analyzed in sections 4.3-5, followed by a discussion of its limitations.

Critical weaknesses

The paper is not very well written:

  • The third contribution "near-OOD signal lies in covariance directions" is oddly stated, since the term "covariance directions" is hardly used afterwards. It should be made clear that it is a result of the g2 detector and its analysis (if i am not mistaken).
  • zz and the feature vector are introduced at the beginning but not reused in the experimental part, where they should be recalled and used.
  • In the same vein: batch data are never clearly defined in the experiments. It should be clearly stated that they follow equation (1) (and if this is not the case, it should be explained).
  • Section 3.3.2 on closed forms is hard to follow due to abrupt references to appendix.
  • S0S_0, in equation (13), is only implicitely defined and it does not directly match the SnS_n in appendix C, which is not defined for n=0n=0.
  • minor remarks on notations: -KL and (J) (for KL may be acceptable) not really defined while they appear in a central proposition; the tensor (?) over z notation in 3.3.3 not even defined in the appendix (and not in official iclr notations).

Even though the idea is appealing, the contribution itself seems limited, as it is mainly an aggregation of well-established works: the first contribution does not seem to be a real one, and one could almost argue that the second and third could be merged.

Questions for authors

Could you report standard deviations, at least for table 1?

About table 2: "the near-OOD gain of g2 therefore comes from the directional term". First, "anisotropic second moment" is not used elsewhere but i suppose it means the third term of equation (13). The scores are equal to g2: maybe more conclusions can be drawn from these results. Did you try to employ this term only as a score for more experiments ? If you have time maybe a few more results can be of use so we can compare the directional term with g2 (or not if you provide a theoretical reason for not using it).

Could you state more clearly your contribution with respect to the works you cite in section 2 ? I understand space is limited in the paper but, fortunately, we have more room here.

6 comments

We thank the reviewer for the careful reading and constructive suggestions on presentation, notation, and experimental reporting.

W1: The third contribution "near-OOD signal lies in covariance directions" is oddly stated, since the term "covariance directions" is hardly used afterwards. It should be made clear that it is a result of the g2 detector and its analysis (if i am not mistaken).

We agree. Contribution 3 in Sec. 1 is now titled "A decomposition of g2g_2 that locates its near-OOD gain" and states the finding as a result of analyzing g2g_2. It no longer uses the term "covariance directions". It instead describes the directions along which the batch covariance departs from the reference, which correspond to the directional term of Eq. (13), the term used throughout Sec. 4.3 (Tables 3–4, Figure 4).

W2 & W3: zz and the feature vector are introduced at the beginning but not reused in the experimental part [...] batch data are never clearly defined in the experiments. It should be clearly stated that they follow equation (1).

We updated Sec. 4.1 to state explicitly that: (1) the feature vector z=ϕ(x)∈Rdz = \phi(x) \in \mathbb{R}^d is the encoder's CLS token (d=768d=768 for ViT-B/16, d=384d=384 for DINOv2), and (2) each evaluation batch follows Eq. (1) by combining (1−π)n(1-\pi)n ID test samples from pp with πn\pi n OOD samples from aa (App. A.2).

W4: Section 3.3.2 on closed forms is hard to follow due to abrupt references to appendix.

We rewrote the opening of Sec. 3.3.2 so that reference whitening and the orthogonality of the degree-1 and degree-2 Hermite shells are explained in the main text before referring to App. B.3.

W5 & W6: S0S_0, in equation (13), is only implicitely defined and it does not directly match the SnS_n in appendix C [...] -KL and (J) [...] not really defined [...] the tensor (?) over z notation in 3.3.3 not even defined.

We fixed all of these notations:

  • Defined the traceless second-moment shift S0:=S−tr⁡SdIS_0 := S - \frac{\operatorname{tr}S}{d}I right before Eq. (13), and renamed the streaming estimates in App. C to S^(n)\widehat{S}^{(n)} and m^(n)\widehat{m}^{(n)} to avoid subscript clashes with S0S_0.
  • Defined KL(a∥p)\mathrm{KL}(a \Vert p) and Jeffreys divergence J(a,p)=KL(a∥p)+KL(p∥a)J(a,p) = \mathrm{KL}(a \Vert p) + \mathrm{KL}(p \Vert a) in Proposition 1, defined the ℓ\ell-fold tensor power z⊗ℓz^{\otimes \ell} in Sec. 3.3.3, and added all of them to the notation table in App. A.1.
  • Also resolved the other notation clashes we found in the rest of the paper.

W7 & Q3: Even though the idea is appealing, the contribution itself seems limited, as it is mainly an aggregation of well-established works: the first contribution does not seem to be a real one, and one could almost argue that the second and third could be merged. [...] Could you state more clearly your contribution with respect to the works you cite in section 2?

We rewrote Contribution 1 in Sec. 1 to clarify what the two-axis formulation brings (linking batch OOD scoring to restricted polynomial readouts and separating batch detection from sample ranking). We keep Contributions 2 and 3 separate because Contribution 2 is methodological (the closed-form detector g2g_2 with O(d2)O(d^2) streaming updates and leave-one-out scoring), whereas Contribution 3 is an empirical finding about representations enabled by decomposing g2g_2.

We also revised Sec. 2 to contrast our work directly with each research thread:

  • vs. Test-time OOD detection (RTL++, OODD, AdaODD): Prior methods use the batch to adapt or pseudo-label an existing ID score. Our readout is defined directly by the batch-to-reference moment comparison, yielding a closed form.
  • vs. Semi-supervised novelty detection (Blanchard et al.): Classical work targets q/pq/p with guarantees for a generic classifier class; we characterize what a finite batch of size nn in dimension dd can resolve by organizing readouts by polynomial degree.
  • vs. General density-ratio estimation (uLSIF, DDE): Kernel and neural estimators set resolution implicitly via model architecture and regularization. Restricting to degree-2 polynomials gives a tuning-free closed form that decomposes into interpretable moment terms.
  • vs. Low-degree polynomial methods: The low-degree likelihood ratio of Hopkins and Kunisky et al. is a single scalar statistic for batch-level hypothesis testing (q≠pq \neq p). We evaluate the projected ratio at each sample for sample-level ranking on the batch itself and show that ranking requires substantially larger batches to estimate the shift direction (Sec. 4.4, App. D.8).

Q1: Could you report standard deviations, at least for table 1?

We added standard deviations across seeds to all entries in Tables 1 and 3.

Q2: About table 2: "the near-OOD gain of g2 therefore comes from the directional term". First, "anisotropic second moment" is not used elsewhere [...] The scores are equal to g2: maybe more conclusions can be drawn from these results. Did you try to employ this term only as a score for more experiments?

  1. Terminology in Table 2: We updated this table (now Table 3) to name each term explicitly by its statistic: mean shift mm, trace tr⁡S\operatorname{tr}S, and traceless shift S0S_0 (12z⊤S0z\frac{1}{2}z^\top S_0 z, the third term of Eq. (13), which "anisotropic second moment" referred to). We also unified the terminology throughout the paper.
  2. Directional term alone: We extended Table 3 to all 44 configurations. The directional term 12z⊤S0z\frac{1}{2}z^\top S_0 z alone matches g2g_2 on near-OOD (0.8530.853 vs. 0.8530.853). We retain the full g2g_2 because (i) the directional term alone is blind to isotropic scale changes (S∝I  ⟹  S0=0S \propto I \implies S_0 = 0), whereas g2g_2 responds to any first- or second-moment shift (Eq. (12)); and (ii) at low contamination g2g_2 outperforms the centered directional readout (π∈{0.005,0.01,0.05}\pi \in \{0.005, 0.01, 0.05\}; App. D.9).
  3. What this equality implies: (a) The near-OOD signal is a change in the shape of the batch covariance, not in its mean or overall scale: the directional term matches g2g_2 after centering the covariance and after also removing the mean direction (0.8350.835 and 0.8340.834 vs. 0.8350.835 for g2g_2; cross-fitted, 8 near-OOD configurations of Table 1; Table 4), and the radial term adds little (Table 3). (b) This signal is spread over many directions: near-OOD needs about 8 eigen-directions of C0C_0 to reach 90% of its normalized AUROC, whereas far-OOD saturates at one (Fig. 4). Together, these suggest that near-OOD images lie among the ID class clusters, where pp is high, but overrepresent some directions there. A reference reader cannot tell along which directions the batch departs from the reference.

Thank you.

Regarding your response on "W7 & Q3", you state in the first contribution: "A two-axis view of OOD readouts."

  1. "Organizing readouts by estimand and resolution explains what each estimand can and cannot reveal" + response: "vs. Semi-supervised novelty detection (Blanchard et al.): Classical work targets..." → What did you really bring to Blanchard et al.? The polynomial analysis is not from you, if I'm not mistaken?

  2. "Links batch-based OOD scoring to classical restricted polynomial readouts" + response: "vs. General density-ratio estimation (uLSIF, DDE): Kernel and neural estimators..." → What is the difference with Kanamori et al.? Is it not a degree-2 polynomial readout also?

Reviewer_Eg6cOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies out-of-distribution detection using an unlabeled test batch and a reference set of normal samples. It argues that comparing the test distribution with the reference can reveal anomalies that appear typical under the reference alone. The authors develop simple polynomial scores based on differences in batch statistics, balancing expressiveness against estimation difficulty. Experiments show that a quadratic score using the batch mean and covariance performs well, with improvements on near-distribution anomalies largely explained by directional covariance changes. The method remains sensitive to batch size, anomaly prevalence, and shifts in otherwise normal data.

Critical strengths
  1. This paper has a clear motivation. The paper explains why a sample can be highly probable under the ID distribution yet belong to an anomalous population. The mixture-model argument provides a principled motivation for using unlabeled batch information. The distinction between detecting a batch shift and ranking individual samples is also valuable.

  2. The proposed method is simple, interpretable, and efficient. The degree-2 score requires only batch moments after reference whitening. Its decomposition into mean, radial, and directional components makes its behavior understandable. Leave-one-out scoring explicitly addresses query self-alignment, and the sufficient statistics permit efficient streaming updates. The experiment results also show the efficiency of the proposed method.

  3. The finding of this paper is interesting. In particular, the experiments show that directional covariance changes account for much of the improvement in near-OOD detection, offering useful insight into the information provided by unlabeled test batches.

Critical weaknesses
  1. The proposed method may be vulnerable to class imbalance within the ID distribution. Rare ID classes may be incorrectly flagged as OOD, particularly when their proportions in the test batch differ from those in the reference set. The authors should evaluate robustness to ID class imbalance and class-prior shifts, including false-positive rates for minority classes.

  2. The advantage of using population-level information for individual-level detection needs clearer justification. Although batch statistics can reveal an overall distributional shift, identifying which individual samples are OOD requires finer discrimination. Why should maximizing a batch–reference contrast produce reliable individual rankings? The paper acknowledges that a positive population mean gap can coexist with poor AUROC, but does not establish conditions under which population-level evidence translates into accurate sample-level detection. Clarifying these conditions would strengthen the motivation for the proposed objective.

Questions for authors

Please see the weaknesses.

2 comments

We thank the reviewer for highlighting the clarity of our motivation, the simplicity and efficiency of the degree-2 readout, and our empirical findings on directional covariance shifts.

W1: The proposed method may be vulnerable to class imbalance within the ID distribution. Rare ID classes may be incorrectly flagged as OOD, particularly when their proportions in the test batch differ from those in the reference set. The authors should evaluate robustness to ID class imbalance and class-prior shifts, including false-positive rates for minority classes.

Eq. (1) assumes that the ID part of the batch follows the reference pp, so a class-prior shift lies outside the model, as covariate shift does (Sec. 4.5). We measured its effect on the 18 configurations of Table 1 (nref=5,000n_{\mathrm{ref}}=5{,}000, n=2,000n=2{,}000, π=0.1\pi=0.1, 10 seeds) with long-tailed ID classes (imbalance ratio ρ=100\rho=100) in the reference, the batch, or both. FPR is the FPR at 95% OOD TPR on the stated third of ID classes, shown as balanced →\to long-tailed.

SettingAUROC: g2g_2 / Mahalanobis++ / kkNNFPR measured onFPR g2g_2FPR Mahalanobis++FPR kkNN
Balanced0.931 / 0.828 / 0.801
Same long tail in ref & batch0.943 / 0.861 / 0.843rarest third (8 CIFAR configurations)0.287 →\to 0.3810.631 →\to 0.8450.599 →\to 0.819
Long tail in reference only0.876 / 0.774 / 0.721rarest third in the reference0.302 →\to 0.5270.515 →\to 0.7610.558 →\to 0.812
Long tail in batch only0.872 / 0.831 / 0.806most frequent third in the batch0.296 →\to 0.4630.504 →\to 0.5000.549 →\to 0.541

The same-long-tail FPR is over the 8 CIFAR configurations; at ρ=100\rho=100 the rarest ImageNet-1K classes have no batch samples.

  • Long tail in reference only: As the reviewer expected, g2g_2 flags classes that are rare in the reference more often (rare-class FPR 0.302→0.5270.302 \to 0.527). Mahalanobis++ and kkNN degrade in the same way.
  • Long tail in batch only: g2g_2 is the only detector whose FPR on the frequent classes rises (0.296→0.4630.296 \to 0.463); on the ten ImageNet-1K configurations it is above Mahalanobis++ (0.4940.494 vs. 0.4080.408).

W2: The advantage of using population-level information for individual-level detection needs clearer justification. [...] Why should maximizing a batch–reference contrast produce reliable individual rankings? The paper [...] does not establish conditions under which population-level evidence translates into accurate sample-level detection.

The link from population-level deflection to individual-level ranking rests on two steps:

  1. Unrestricted optimum: Under q=(1−π)p+πaq = (1-\pi)p + \pi a, the posterior probability Pr⁡(ood∣z)\Pr(\mathrm{ood}\mid z) is strictly increasing in the density ratio r(z)=q(z)/p(z)r(z) = q(z)/p(z) (Eq. (2)). Without degree restrictions, maximizing deflection recovers r−1r-1 up to scale (Sec. 3.3.1), which yields the Bayes-optimal individual ranking.
  2. Finite-degree projection gkg_k: Because ROC depends only on score ordering, gkg_k preserves the Bayes-optimal ranking whenever it is a strictly increasing function of rr. This holds when r−1∈Vkr - 1 \in V_k (Theorem 1(ii)), as well as for Gaussian mean shifts (g1g_1) and subspace covariance shifts (g2g_2), where g1g_1 and g2g_2 match the Neyman–Pearson AUROC on synthetic data (0.9830.983 and 0.8970.897; App. D.8).

Outside these cases, ranking degrades with the residual of r−1r-1 orthogonal to VkV_k: positive deflection guarantees that OOD points receive a higher average score than ID points, though not pointwise ordering (counterexample in App. B.4). We added a summary of this argument right after Theorem 1.

Thank you for the detailed response, which has addressed most of my concerns. I will therefore maintain my positive rating of 3.

Reviewer_fjtrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors investigates OOD detection based on an unlabelled test batch and design a polunomial approximation to the batch-to-reference desntiy ratios. They disign a practical implementation by leveraging the gap between the first and second moment uder a gaussian-based model.

Critical strengths
  1. this paper is well written.

  2. figures are informative.

  3. experiemnts are conducted on a wide range of backbones

  4. the motivation of this paper is appealing.

Critical weaknesses
  1. While I agree that whitening can fix first and second moments; it can not necessarily make the third- and fourth-moment identities be the Gaussian polynomial Gram matrix. This makes eq. (13) theorectically unjustified, since it would differ from deflection-optimal quadratic readout for the reference distribution.

  2. The authors only compare their method with a limited number of baselines, not to mentioning that the selected baselines are out-dataed. Why the authors only use SVHN and C10(C100) as OOD datasets for evaluating performance on C100 (C10) as ID datasets.

  3. Just as explcitly acknowledged in Remark 1, the key results in this paper largely results from classical likelihood-ratio optimality, projection arguments, and Hermite expansions. It is unclear regarding the novelty of this paper.

  4. The use of n≍dk/2n\asymp d^{k/2} to explain practical viability of different degrees is theorectically under-supported.

Questions for authors

see weakness.

1 comment

We thank the reviewer for the positive assessment of our paper's writing, figures, motivation, and backbone coverage, as well as the thoughtful questions below.

W1: While I agree that whitening can fix first and second moments; it can not necessarily make the third- and fourth-moment identities be the Gaussian polynomial Gram matrix. This makes eq. (13) theorectically unjustified, since it would differ from deflection-optimal quadratic readout for the reference distribution.

We agree that whitening does not make pp Gaussian (App. A.3). For a general whitened pp, the deflection-optimal quadratic readout g2⋆g_2^\star uses a Gram matrix containing the third and fourth moments of pp, whereas g2g_2 in Eq. (13) uses the identity Gram matrix of the Gaussian basis. Thus, g2g_2 is a plug-in Gram surrogate for g2⋆g_2^\star that avoids estimating noisy fourth-order reference moments in dd dimensions while retaining two key properties:

  1. Same detection support: Both g2g_2 and g2⋆g_2^\star are driven by the same contrast vector (m,S)(m, S) and vanish under the exact same condition (m=0,S=0m=0, S=0). Whenever g2⋆g_2^\star achieves positive deflection, g2g_2 does as well (Eq. (12)).
  2. Comparable finite-sample AUROC: A quadratic least-squares fit with the empirical reference Gram matrix and a ridge selected on held-out data yields a near-OOD AUROC of 0.83500.8350, compared to 0.83520.8352 for g2g_2 (cross-fitted, 8 near-OOD configurations of Table 1). Weaker regularization, which is closer to g2⋆g_2^\star, does not improve on g2g_2 (0.8270.827 to 0.8340.834; App. D.6).

We now state this surrogate interpretation explicitly in Sec. 3.3.2 and App. B.3; the empirical comparison is in Sec. 4.3 and App. D.6.

W2: The authors only compare their method with a limited number of baselines, not to mentioning that the selected baselines are out-dataed. Why the authors only use SVHN and C10(C100) as OOD datasets for evaluating performance on C100 (C10) as ID datasets.

The main comparison is run on the full OpenOOD v1.5 benchmark (44 configurations), covering CIFAR-10, CIFAR-100, ImageNet-200, and ImageNet-1K as ID datasets and all standard near- and far-OOD datasets (TinyImageNet, MNIST, Textures, Places365, SSB-hard, NINCO, iNaturalist, OpenImage-O), alongside recent baselines (Mahalanobis++, 2025) and test-time methods (RTL++, 2024; OODD, 2025; and the newly added POT, 2025, and AdaODD, 2026). In the submitted version, Table 1 showed a subset for space reasons and the full grid was summarized in the appendix. Across all 44 configurations, g2g_2 has mean AUROC 0.9260.926 on all OOD and 0.8530.853 on near-OOD, the highest among the methods that use no ID labels (0.9000.900 and 0.8070.807 for DDE, the next best). POT, which uses ID class labels at detection time, is slightly above g2g_2 on both (0.9280.928 and 0.8550.855), within one standard deviation, and also on far-OOD (0.9700.970 vs. 0.9670.967) and in FPR@95 (0.2730.273 vs. 0.3030.303; App. D.3).

We updated Sec. 4.1 and the Table 1 caption to highlight the full benchmark scope upfront, added a summary of all 16 methods in Table 2, and added complete per-configuration tables in App. D.3.

W3: Just as explcitly acknowledged in Remark 1, the key results in this paper largely results from classical likelihood-ratio optimality, projection arguments, and Hermite expansions. It is unclear regarding the novelty of this paper.

We agree that the mathematical building blocks are classical, and we do not claim a new optimality theorem. Our contribution lies in formulating label-free batch OOD detection through finite-degree density-ratio projections, yielding two new findings for OOD detection:

  1. A tuning-free closed-form detector (g2g_2): Projecting onto degree-2 polynomials gives a closed-form readout requiring no hyperparameter tuning or iterative optimization.
  2. Where the near-OOD signal lives: Decomposing g2g_2 reveals that the near-OOD gain in evaluated encoders comes specifically from directional (traceless) covariance shifts (S0S_0), rather than mean shifts or isotropic variance changes (Tables 3–4).

We revised Sec. 1 and Sec. 2 to state these points clearly.

W4: The use of n≍dk/2n\asymp d^{k/2} to explain practical viability of different degrees is theorectically under-supported.

We agree that n≍dk/2n \asymp d^{k/2} is a heuristic scaling rule rather than a formal sample-complexity bound (noted in Sec. 3.3.3 and App. B.5), used only as an intuitive scale for the O(dk)O(d^k) degrees of freedom in a degree-kk shell. Empirically, fitted scaling exponents on synthetic data are 0.57,1.03,1.410.57, 1.03, 1.41 for k=1,2,3k=1, 2, 3 (App. D.8), and on real benchmarks (d≈384–768,n=2,000d \approx 384\text{--}768, n=2{,}000), where d3/2≫nd^{3/2} \gg n, the centered degree-3 readout improves on g2g_2 in only 1 of 18 configurations, by less than the seed variation (App. D.8).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their constructive feedback. Revisions in the manuscript are highlighted in blue (minor notation cleanups are left in black). Below is a summary of the main updates:

  • OpenOOD v1.5 preprocessing (Sec. 4, Table 1, App. D.1, D.3): Re-extracted all features with the exact OpenOOD v1.5 test pipeline and recomputed all results; the main conclusions are unchanged. Stated the full 44-configuration scope upfront in Sec. 4.1 and the Table 1 caption.
  • Non-Gaussian reference (Sec. 3.3.2, App. B.3): Stated that g2g_2 is a Gram surrogate for the deflection-optimal quadratic readout g2⋆g_2^\star when the whitened reference pp is non-Gaussian, and how the two are related.
  • Baselines and reporting (Sec. 4.2, Tables 1–3, App. D.2, D.3): Added POT and AdaODD, a summary of all 16 methods on the full grid (Table 2), seed standard deviations (Tables 1 and 3), and per-configuration tables (App. D.3). Among the methods that use no ID labels, g2g_2 has the highest mean AUROC (0.9260.926; next best DDE, 0.9000.900). POT, which uses ID class labels, is slightly above g2g_2 (0.9280.928), within one standard deviation.
  • Positioning (Sec. 1, Sec. 2): Rewrote Contributions 1 and 3 and stated in Sec. 2 how our readouts differ from each line of related work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.