Pith. sign in

REVIEW 4 major objections 5 minor 50 references

Zero-shot OOD detector rankings reverse across domains and models, so benchmark winners are not deployment winners.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 12:52 UTC pith:JQGRYE5H

load-bearing objection Solid negative audit of OOD detector ranking transfer, but the CEG fix is developed and evaluated on the same tasks — treat the positive claim as promising but unproven. the 4 major comments →

arxiv 2607.26582 v1 pith:JQGRYE5H submitted 2026-07-29 cs.CV

Level, Sharpness, and Corpus: Why Zero-Shot OOD Detector Rankings Do Not Transfer

Ignacio M. De la Jara , Cristian Rodriguez-Opazo , Stephen Gould , Damith Ranasinghe This is my paper
classification cs.CV
keywords zero-shot OOD detectionvision-language modelsdetector portabilitycomplementary evidencenon-compensatory fusionlevel and sharpnessFPR95CLIP
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that the common practice of choosing a zero-shot OOD detector by benchmark ranking is unsafe: rankings reverse when the in-distribution dataset or the underlying vision-language model changes, and every detector exceeds 80% FPR95 on at least one domain. It explains these reversals by showing that detectors rely on complementary evidence channels — absolute match level, relative or spatial sharpness, and external semantic coverage — and that no single channel is universally informative. A simple proposition shows that level and sharpness cannot generally be recovered from one another, which is why no fixed detector transfers reliably. The paper then proposes the Complementary Evidence Guard (CEG), a detector-agnostic wrapper that preserves all channels by taking the minimum of their empirical in-distribution percentiles, and reports that this reduces FPR95 for every wrapped detector while shrinking the spread between detectors.

Core claim

The central claim is that zero-shot OOD detector rankings are not portable across deployments. Through a controlled audit across seventeen in-distribution datasets, three vision-language models, and seven detectors, the paper shows that the best detector changes with both the ID domain and the VLM, and that no detector stays below 80% FPR95 everywhere. The explanation is that detectors preserve different evidence: softmax-based scores like MCM discard absolute level and keep only sharpness, energy and MaxLogit keep level, and WordNet-based methods depend on external corpus coverage. Proposition 1 shows MCM is shift-invariant and approximates the centred maximum, while Energy preserves level

What carries the argument

The central objects are the complementary evidence channels extracted from CLIP-style cosine logits: level (the maximum cosine match to ID anchors) and sharpness (how much the best match stands out from the mean, plus a spatial sharpness term). Proposition 1 identifies MCM as a shift-invariant approximation to the centred max, preserving only sharpness, while the log-sum-exp decomposition shows Energy preserves level to leading order. CEG maps the base score, level, and sharpness each to empirical in-distribution percentiles and then takes their minimum, enforcing a non-compensatory veto: any single atypical channel rejects the sample. The minimum is the load-bearing mechanism that directly

Load-bearing premise

The load-bearing premise is that CEG's hyperparameters and the level/sharpness channel definitions, tuned on capped versions of the same five strict tasks, transfer to new ID domains and VLMs; the paper's own evidence tiers show that most headline results come from development/evaluation overlap, with only a handful of genuinely prospective tests.

What would settle it

Run CEG with frozen hyperparameters on a new, non-ImageNet-like ID domain (for example, medical imaging or aerial scenes) using CLIP features and show that the family-balanced FPR95 does not improve over the best raw base detector; or demonstrate that a detector explicitly preserving both level and sharpness through learned fusion outperforms CEG on a prospective benchmark, contradicting the claim that the minimum rule's gains come from evidence preservation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Benchmark averages alone are insufficient for detector selection; worst-domain risk and revalidation across ID domains and VLMs should be reported.
  • Detector rankings should be re-checked whenever the in-distribution data or the underlying vision-language model changes, since both flip the preferred method.
  • Preserving complementary evidence rather than committing to a single score is a more robust deployment strategy, with CEG improving every wrapped detector without OOD samples or auxiliary corpora.
  • The vocabulary-size confound (growing K alone changes detector rankings) implies that benchmark comparisons should control for label-set size.
  • CEG reduces sensitivity to detector choice, but it does not crown a universal winner; the best guarded detector still shifts across evaluation weightings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The level/sharpness decomposition likely applies beyond vision-language models to any logit-based classifier, so a similar non-compensatory guard could stabilize OOD detection in standard deep networks.
  • The strong dependence of WordNet-based methods on corpus coverage suggests that future external-semantic detectors should mine or adapt the negative vocabulary to the target ID domain rather than reuse a fixed pool.
  • The paper's own appendix flags that most headline CEG results were developed and evaluated on capped versions of the same five tasks; the only truly prospective checks are the ImageNet-10/20 reciprocal test and the SigLIP2/PE port, several cells of which regress, so the transferability claim is weaker than the headline numbers suggest.
  • The non-compensatory veto principle could be useful in other safety-critical fusion settings where distinct cues fail independently and a single atypical channel should be enough to raise an alarm.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper investigates whether zero-shot OOD detector rankings transfer across deployment settings. Across seventeen ID datasets, three VLMs, and seven detectors, the authors find that rankings reverse with ID domain and VLM, that every detector exceeds 80% FPR95 on at least one domain, and that the preferred detector depends on both the ID data and the encoder. They attribute these reversals to complementary evidence channels: absolute match level, relative/spatial sharpness, and, for WordNet-based methods, external semantic coverage. A proposition shows that MCM discards level and approximates sharpness, while Energy preserves level, explaining complementary failures. The authors then propose CEG, a wrapper combining a base detector with level and sharpness channels through an empirical-percentile minimum fusion. On five strict OpenOOD tasks, CEG lowers family-balanced FPR95 for all wrapped detectors, and a broader audit shows contraction in detector spread. The manuscript is unusual in that it explicitly labels evidence tiers and discloses development/evaluation overlap in the appendices.

Significance. If the negative portability finding is accepted, it is a timely and important caution for the zero-shot OOD detection community: benchmark rankings on ImageNet-centred suites do not transfer across ID domains or VLM families. The paper's strengths include a controlled audit with frozen implementations, paired resampling intervals, a benchmark-composition probe, and an unusually transparent evidence-tier disclosure. The positive CEG proposal is well motivated by the evidence-channel diagnosis and the channel-substitution controls are informative; however, as discussed below, the deployment-robustness claim for CEG is not yet established by held-out task evidence. The paper's contribution is therefore strongest as a diagnosis/audit and weaker as a validated new method.

major comments (4)
  1. [§5 / Appendix B (Evidence tiers)] The headline CEG results in Table 1a are measured on the same five strict task domains whose capped versions were used to develop the channel forms, k=10, Q0.9, 3x3 kernel, and minimum fusion. As Appendix B states, the uncapped rerun is sampling-level stability evidence, not new-task confirmation. The only prospective checks are the reciprocal ImageNet-10/20 experiment (still ImageNet, hence same domain and correlated subsets) and the SigLIP2/PE port. The port shows five of fifteen point estimates worsen, including both SigLIP2 ImageNet tasks and PE CIFAR-100 unresolved, and CEG(MaxLogit) worsens the SigLIP2 aggregate. Thus the claim in the abstract that CEG 'reduces detector sensitivity' across deployments rests on development/evaluation overlap rather than on held-out task evidence. The claim should be narrowed or supported by a genuinely out-of-development evaluation (e.g., new ID dat
  2. [Appendix E (Protected-weight provenance)] The 2:1 shrinkage ratio for CEG-P was selected after inspecting the transfer failures it was designed to mitigate, and the appendix correctly labels this as post-result development evidence. However, Section 5 and Figure 7 present CEG-P as a protected variant and report broad all-domain improvements, and Table F1 is titled 'protected audit.' Because the weight was chosen after observing the failures, all CEG-P comparisons in Tables F1-F4 are circular for that variant. The paper should either clearly present CEG-P as development evidence in the main text, or provide a prospective validation of the protected weight on tasks not used in its selection.
  3. [§5 / Appendix F (Broad deployment audit)] The broad CEG-P audit claims improvements in all 15 near/far deployment summaries and that every interval excludes zero. But the underlying point estimates include multiple regressions, and the intervals are described as nominal and unadjusted in Appendix B. The symmetric fold-wise control is a good sanity check, but it does not address the more serious problem that the audit is post-result development evidence. The strong language 'deployment robustness' and 'reduces deployment sensitivity' is not warranted by the disclosed evidence tier. I would ask the authors to separate confirmatory results from descriptive development evidence in the main-text claims.
  4. [Proposition 1 / §4] The expansion for MCM at large temperature is correct and helpful, but the claim that MCM 'preserves only the sharpness channel' should be stated with the caveat that the expansion is asymptotic and holds for T >> R; the empirical approximation is verified in the appendix, but the main-text phrasing overstates exactness. This is not a blocking issue, but it affects the interpretive precision of the mechanism argument.
minor comments (5)
  1. [Table B1] Several rows appear malformed or have misaligned column counts (e.g., the CUB-200 row). Please check the table rendering and ensure each detector has a value in every row.
  2. [Eq. (1)] The notation z_ID for standardization is used before the mean/standard deviation are defined; state explicitly that z_ID denotes (value - mean)/std estimated on the unlabeled ID calibration set.
  3. [Main text] 'Family-balanced' is used in the abstract and Section 5 but defined only in the appendix; give a one-sentence definition in the main text.
  4. [Figure 2 caption] 'All-five winner' and 'frozen global subset' are referenced in the caption before these terms are defined in the text; define or rephrase.
  5. [Appendix F] The text reports 896/1,185 point improvements for CEG-P; given the large number of comparisons, a multiplicity correction or explicit statement that these are unadjusted descriptive counts would be useful.

Circularity Check

2 steps flagged

CEG's headline improvements are partly self-referential due to development/evaluation overlap; the portability audit and Proposition 1 are independent.

specific steps
  1. fitted input called prediction [Section 5 'Evaluation protocol'; Appendix B 'Evidence tiers']
    "CEGisdevelopedonlyoncapped versions of these tasks, whereas the reported results use the full uncapped benchmarks. ... The channel forms andk=10,Q 0.9,3×3 kernel, and minimum fusion were developed on capped versions of the same five strict task domains; their uncapped rerun is sampling-level stability evidence, not new-task confirmation."

    The headline family-balanced gains (GL-MCM 38.1→28.8, MCM 42.6→30.5) are measured on ImageNet-1K/200, CIFAR-10/100, and MNIST — the same task domains used to select the channel forms, k, Q, kernel size, and minimum fusion rule. The uncapped rerun changes image samples but not task domains, so it is stability evidence rather than a confirmatory prediction. Calling the strict-task results evidence that CEG reduces deployment sensitivity is therefore partly self-referential for those tasks. The only out-of-development tests the paper identifies are the reciprocal ImageNet-10/20 check and the SigLIP2/PE port, and the port shows five of fifteen point estimates worsen, with PE CIFAR-100 unresolved.

  2. fitted input called prediction [Appendix E 'Protected-weight provenance'; Section 5 'Broad deployment audit']
    "The2:1ratio was chosen after inspecting the transfer failures it was designed to mitigate; it is not a held-out hyperparameter choice, and noCEG-Pcomparison is presented as confirmation."

    CEG-P is the variant used for the broad deployment-robustness claim ('Evaluating CEG-P ... far OOD improves in all 15 near/far deployment summaries and both near and far improve simultaneously in 14'). Because its λ=1/3 protection weight was selected after inspecting the transfer failures it mitigates, the broad-audit success of CEG-P is not an out-of-sample validation of the protection mechanism. The paper explicitly labels this post-result development evidence, but the deployment-robustness claim for the protected variant is still partly defined by the failures it was built to fix.

full rationale

The central negative finding — zero-shot OOD detector rankings reverse across seventeen ID datasets and three VLMs — is well supported by the audit and does not depend on CEG. Proposition 1 is a proved mathematical identity (MCM is shift-invariant and approximates the centred maximum), not a circular input. The circularity concern is confined to the positive CEG claim: the strict five-task headline was developed on capped versions of the same five task domains, and the CEG-P protection weight was chosen after observing the transfer failures it mitigates. The paper is unusually transparent about this, explicitly labeling the uncapped rerun as 'sampling-level stability evidence' and CEG-P as 'post-result development evidence,' which prevents the issue from being concealed. Nevertheless, the headline 'reduces deployment sensitivity' rests partly on development/evaluation overlap; the genuinely prospective checks are limited and show mixed regressions. The one self-citation (M. De la Jara et al. 2025) concerns a local-feature implementation detail and is not load-bearing. Overall, the negative audit and the mathematical decomposition stand independently, so the paper is not globally circular, but the positive deployment-robustness claim is materially weakened by its self-referential evidence tier.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 1 invented entities

CEG introduces four hand-chosen modeling ingredients (k, kernel, lambda, min-rule), all disclosed but several selected on the same task family used for evaluation. The conceptual level/sharpness channels are the paper's own interpretive framework, not independently measured quantities. The main empirical audit, in contrast, relies mostly on standard benchmark assumptions and published detector definitions.

free parameters (4)
  • k (number of local patches for level channel) = 10
    Chosen during development on capped strict tasks; controls how many high-confidence patches contribute to the local level term.
  • spatial sharpness kernel (Q0.9, 3x3) = Q0.9, 3x3
    Developed on capped strict tasks; defines the spatial sharpness channel in Eq. 1.
  • CEG-P shrinkage weight lambda = 1/3
    Explicitly selected after inspecting the transfer failures it was designed to mitigate; the paper states it is not a held-out hyperparameter choice.
  • minimum fusion rule = min
    The non-compensatory minimum was retained over Simes and Fisher, which achieve lower error on the development suite; this is a design choice with measurable numerical consequences, not forced by the data.
axioms (6)
  • standard math Softmax shift-invariance and log-sum-exp identity
    Used in Proposition 1 to show MCM discards absolute level and Energy preserves it to leading order; invoked in Section 4.
  • domain assumption CLIP logits encode two complementary signals: absolute level and relative/spatial sharpness
    Central diagnostic premise; supported empirically by Figure 4 and channel ablations, but treated as the explanatory framework throughout Section 4.
  • domain assumption Unlabeled ID calibration (1,000 images) suffices to estimate empirical percentiles and the operating threshold
    CEG depends on this throughout; the paper explicitly claims no finite-sample coverage guarantee because channels are dependent.
  • domain assumption OpenOOD-v1.5 OOD suites are a valid, clean evaluation base
    Motivated by the duplicate/proximity analysis, but the paper ultimately adopts OpenOOD-v1.5 as the evaluation ground truth for the main claims.
  • domain assumption The seven detectors are representative of zero-shot OOD detection
    The portability conclusion is scoped to this detector family; representativeness is asserted rather than proven.
  • domain assumption WordNet coverage, not the mining procedure, limits NegLabel/CSP transfer
    Inferred from 89.5-93% vocabulary overlap across domains; alternative explanations such as class-name mismatch or image-domain shift are not fully controlled.
invented entities (1)
  • Complementary evidence channels: level L and sharpness S no independent evidence
    purpose: Explain ranking reversals and define the CEG fusion rule
    The channels are defined from the same VLM logits and evaluated on the same task family used to build CEG; the prospective VLM port is internal evidence, not an external falsifiable handle.

pith-pipeline@v1.3.0-daily-deepseek · 33216 in / 14034 out tokens · 133151 ms · 2026-08-01T12:52:31.559423+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Level, Sharpness, and Corpus: Why Zero-Shot OOD Detector Rankings Do Not Transfer." pith.science (2026). https://pith.science/paper/JQGRYE5H

@misc{pith2026260726582,
  author       = {Pith},
  title        = {Pith review of: Level, Sharpness, and Corpus: Why Zero-Shot OOD Detector Rankings Do Not Transfer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JQGRYE5H}},
  note         = {Machine review of arXiv:2607.26582}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Selecting a zero-shot out-of-distribution (OOD) detector for a new deployment is typically based on benchmark rankings, implicitly assuming that the highest-ranked detector will transfer across domains. We show that this assumption does not hold. Through a controlled portability audit across seventeen in-distribution datasets, three vision-language models, and seven representative zero-shot OOD detectors, we find that detector rankings reverse across deployments, every detector exceeds $80\%$ FPR95 on at least one domain, and the preferred detector depends on both the in-distribution data and the underlying VLM. We trace these reversals to complementary evidence channels in vision-language logits. Corpus-free detectors rely on different combinations of absolute match level and relative or spatial sharpness, while WordNet-based methods additionally depend on external semantic coverage. A simple proposition shows that level and sharpness cannot generally be recovered from one another, explaining why no single detector transfers reliably across deployments. Motivated by this diagnosis, we introduce the Complementary Evidence Guard (CEG), a detector-agnostic wrapper that preserves complementary evidence through a non-compensatory fusion of the base detector, level, and sharpness using only empirical in-distribution percentiles. Controls replacing these channels with entropy, logit variance, or random noise do not reproduce the gains. Without OOD samples, auxiliary corpora, or learned fusion, CEG reduces detector sensitivity and improves GL-MCM from $38.1$ to $28.8$ and MCM from $42.6$ to $30.5$ family-balanced FPR95.

Figures

Figures reproduced from arXiv: 2607.26582 by Cristian Rodriguez-Opazo, Damith Ranasinghe, Ignacio M. De la Jara, Stephen Gould.

Figure 1
Figure 1. Figure 1: Benchmark winners do not transfer across ID domains. (a) Detector rankings reverse across ID domains. (b) Different ID domains favour different evidence. (c) Every detector fails on at least one ID domain, showing that no single zero-shot score is universally reliable. themselves remain valid. What fails is the assumption that benchmark rankings transfer across deployments. This assumption has become wides… view at source ↗
Figure 2
Figure 2. Figure 2: Portability fails because detectors preserve different evidence. (a) The best corpus-free detector changes across ID domains. (b) The contribution of external semantic coverage depends on the ID domain and VLM. (c) The cost of fixing one score: filled markers show the average FPR95 gap to the best score for each unit, and open markers show the worst-case gap. Farther right means worse transfer. not control… view at source ↗
Figure 3
Figure 3. Figure 3: Benchmark composition is part of the audit. (a) A verified Places/ImageNet duplicate and a selected SUN neighbour illustrate that label-defined OOD need not be vi￾sually disjoint. (b) Maximum-cosine distributions overlap substantially between ImageNet and the traditional OOD sources. Full protocol: Appendix C. universally better detectors; they reward different evidence channels. This naturally raises the … view at source ↗
Figure 4
Figure 4. Figure 4: Level and sharpness separate different shifts. Sharpness separates ImageNet, whereas absolute match level separates CUB-200, explaining the MCM/MaxLogit rever￾sal. Contours and header FPR95 use a diagnostic subset (20 ID images/class; at most 3,000/OOD source); Tables 1 and B1 report separate protocols. Horizontal position is mean cosine, the level term in MaxLogit. lap between the corresponding maximum-co… view at source ↗
Figure 5
Figure 5. Figure 5: Label-set construction changes the detector. (a) Mean FPR95 on a fixed suite under ImageNet class sub￾sampling (eight draws; bands ±1 s.d.). (b) Other global￾logit scores minus MCM; positive favours MCM, and MSP=MCM under this scoring convention. −1 0 1 2 3 negative pressure (ID-IQR) → ImageNet-1K CIFAR-100 MNIST 86 / 47 67 / 87 77 / 65 AUROC / FPR95 a Negative-label evidence weakens off ImageNet ID OOD b … view at source ↗
Figure 6
Figure 6. Figure 6: External semantic evidence depends on corpus coverage. (a) Negative-label evidence separates ImageNet but weakens off ImageNet. (b) NegMining selects nearly the same WordNet vocabulary across deployments, yet detector performance reverses. Corpus coverage, not vocabulary se￾lection, limits transfer. sharpness-separable domains (ImageNet-1K/200 and Cal￾tech) reward MCM and GL-MCM. Domains where nei￾ther cha… view at source ↗
Figure 8
Figure 8. Figure 8: Uninformative channels cost the minimum. CLIP-B/16 MCM family-balanced FPR95 on five strict tasks and fixed ID-only manifests. Noise is seeded random; en￾tropy and variance independently round to 47.2. Shading is worse than raw. present in MCM, while replacing the best-anchor level sig￾nal with mean affinity improves only to 35.6, showing that the useful level evidence is carried by the best-anchor match r… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 3 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , year =

    A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks , author =. Advances in Neural Information Processing Systems , year =

  2. [2]

    Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li , booktitle =

  3. [3]

    Learning Multiple Layers of Features from Tiny Images , author =

  4. [4]

    Proceedings of the IEEE , volume =

    Gradient-Based Learning Applied to Document Recognition , author =. Proceedings of the IEEE , volume =

  5. [5]

    Wah, Catherine and Branson, Steve and Welinder, Peter and Perona, Pietro and Belongie, Serge , institution =. The

  6. [6]

    European Conference on Computer Vision , year =

    Food-101 -- Mining Discriminative Components with Random Forests , author =. European Conference on Computer Vision , year =

  7. [7]

    Learning Generative Visual Models from Few Training Examples: An Incremental

    Fei-Fei, Li and Fergus, Rob and Perona, Pietro , booktitle =. Learning Generative Visual Models from Few Training Examples: An Incremental

  8. [8]

    IEEE Conference on Computer Vision and Pattern Recognition , year =

    Cats and Dogs , author =. IEEE Conference on Computer Vision and Pattern Recognition , year =

  9. [9]

    Helber, Patrick and Bischke, Benjamin and Dengel, Andreas and Borth, Damian , journal =

  10. [10]

    Everingham, Mark and Van Gool, Luc and Williams, Christopher K. I. and Winn, John and Zisserman, Andrew , journal =. The Pascal Visual Object Classes (

  11. [11]

    IEEE/CVF International Conference on Computer Vision , year =

    Moment Matching for Multi-Source Domain Adaptation , author =. IEEE/CVF International Conference on Computer Vision , year =

  12. [12]

    International Conference on Learning Representations , year =

    Open-Set Recognition: A Good Closed-Set Classifier Is All You Need? , author =. International Conference on Learning Representations , year =

  13. [13]

    In or Out?

    Bitterwolf, Julian and M. In or Out?. International Conference on Machine Learning , year =

  14. [14]

    Wang, Haoqi and Li, Zhizhong and Feng, Litong and Zhang, Wayne , booktitle =

  15. [15]

    and Oliva, Aude and Torralba, Antonio , booktitle =

    Xiao, Jianxiong and Hays, James and Ehinger, Krista A. and Oliva, Aude and Torralba, Antonio , booktitle =

  16. [16]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Places: A 10 Million Image Database for Scene Recognition , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

  17. [17]

    International Conference on Machine Learning , year =

    A Close Look at Negative Label Guided Out-of-Distribution Detection in Pre-trained Vision-Language Models , author =. International Conference on Machine Learning , year =

  18. [18]

    Zhang, Yabin and Zhang, Lei , booktitle =

  19. [19]

    Extract Free Dense Labels from

    Zhou, Chong and Loy, Chen Change and Dai, Bo , booktitle =. Extract Free Dense Labels from

  20. [20]

    Conjugated Semantic Pool Improves

    Chen, Mengyuan and Gao, Junyu and Xu, Changsheng , booktitle =. Conjugated Semantic Pool Improves

  21. [21]

    Advances in Neural Information Processing Systems , volume =

    Revisiting Logit Distributions for Reliable Out-of-Distribution Detection , author =. Advances in Neural Information Processing Systems , volume =

  22. [22]

    arXiv preprint arXiv:2502.14786 , year =

    Tschannen, Michael and Gritsenko, Alexey and Wang, Xiao and Naeem, Muhammad Ferjad and Alabdulmohsin, Ibrahim and Parthasarathy, Nikhil and Evans, Talfan and Beyer, Lucas and Xia, Ye and Mustafa, Basil and H. arXiv preprint arXiv:2502.14786 , year =

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2504.13181 , year =

    Perception Encoder: The Best Visual Embeddings Are Not at the Output of the Network , author =. arXiv preprint arXiv:2504.13181 , year =

    Pith/arXiv arXiv

  24. [24]

    International Conference on Learning Representations , year =

    A Statistical Framework for Efficient Out of Distribution Detection in Deep Neural Networks , author =. International Conference on Learning Representations , year =

  25. [25]

    The Annals of Statistics , volume =

    Testing for Outliers with Conformal p-values , author =. The Annals of Statistics , volume =

  26. [26]

    Kaur, Ramneet and Jha, Susmit and Roy, Anirban and Park, Sangdon and Dobriban, Edgar and Sokolsky, Oleg and Lee, Insup , booktitle =

  27. [27]

    Advances in Neural Information Processing Systems , volume =

    Exploring the Limits of Out-of-Distribution Detection , author =. Advances in Neural Information Processing Systems , volume =

  28. [28]

    Zero-Shot Out-of-Distribution Detection Based on the Pre-trained Model

    Esmaeilpour, Seyed and Liu, Bing and Robertson, Eric and Shu, Lei , booktitle =. Zero-Shot Out-of-Distribution Detection Based on the Pre-trained Model

  29. [29]

    Wang, Hualiang and Li, Yi and Yao, Huifeng and Li, Xiaomeng , booktitle =

  30. [30]

    Li, Tianqi and Pang, Guansong and Bai, Xiao and Miao, Wenjun and Zheng, Jin , booktitle =

  31. [31]

    ICML , year=

    Learning Transferable Visual Models From Natural Language Supervision , author=. ICML , year=

  32. [32]

    NeurIPS , year=

    Delving into Out-of-Distribution Detection with Vision-Language Representations , author=. NeurIPS , year=

  33. [33]

    ICLR , year=

    Negative Label Guided OOD Detection with Pretrained Vision-Language Models , author=. ICLR , year=

  34. [34]

    NeurIPS , year=

    LoCoOp: Few-Shot Out-of-Distribution Detection via Prompt Learning , author=. NeurIPS , year=

  35. [35]

    IJCV , year=

    Zero-Shot In-Distribution Detection in Multi-Object Settings Using Vision-Language Foundation Models , author=. IJCV , year=

  36. [36]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Mysteries of the Deep: Role of Intermediate Representations in Out-of-Distribution Detection , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =. 2510.05782 , archivePrefix =

    arXiv

  37. [37]

    CVPR , year=

    MOS: Towards Scaling Out-of-distribution Detection for Large Semantic Space , author=. CVPR , year=

  38. [38]

    NeurIPS Datasets and Benchmarks , year=

    OpenOOD: Benchmarking Generalized Out-of-Distribution Detection , author=. NeurIPS Datasets and Benchmarks , year=

  39. [39]

    arXiv:2306.09301 , year=

    OpenOOD v1.5: Enhanced Benchmark for Out-of-Distribution Detection , author=. arXiv:2306.09301 , year=

    Pith/arXiv arXiv

  40. [40]

    ICML , year=

    Scaling Out-of-Distribution Detection for Real-World Settings , author=. ICML , year=

  41. [41]

    NeurIPS , year=

    Energy-based Out-of-distribution Detection , author=. NeurIPS , year=

  42. [42]

    ICLR , year=

    A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks , author=. ICLR , year=

  43. [43]

    CVPR , year=

    Describing Textures in the Wild , author=. CVPR , year=

  44. [44]

    CVPR , year=

    The iNaturalist Species Classification and Detection Dataset , author=. CVPR , year=

  45. [45]

    Communications of the ACM , year=

    WordNet: A Lexical Database for English , author=. Communications of the ACM , year=

  46. [46]

    International Journal of Computer Vision (IJCV) , volume=

    Learning to Prompt for Vision-Language Models , author=. International Journal of Computer Vision (IJCV) , volume=

  47. [47]

    CVPR , year=

    Conditional Prompt Learning for Vision-Language Models , author=. CVPR , year=

  48. [48]

    Zhang, Yabin and Zhu, Wenjie and He, Chenhang and Zhang, Lei , booktitle=

  49. [49]

    ICML , year=

    Envisioning Outlier Exposure by Large Language Models for Out-of-Distribution Detection , author=. ICML , year=

  50. [50]

    Transactions on Machine Learning Research (TMLR) , year=

    Generalized Out-of-Distribution Detection and Beyond in Vision Language Model Era: A Survey , author=. Transactions on Machine Learning Research (TMLR) , year=