Pith. sign in

REVIEW 3 major objections 6 minor 92 references

Knowledge Regularized Negative Feature Tuning of Vision-Language Models for Out-of-Distribution Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Scaling and shifting pre-trained text features—rather than learning prompts—lets a vision-language model detect out-of-distribution images on both trained and never-seen classes.

desk verdict A solid, incremental CLIP OOD-tuning method with broad experiments; the unseen-class generalization claim needs a negative-label overlap check before I'd trust the headline number. read the letter →

arxiv 2507.19847 v2 pith:RMHGM2U2 submitted 2025-07-26 cs.CV

classification cs.CV
keywords out-of-distributiondetectionvision-languagemodelsCLIPnegativeprompttuningfeatureknowledgeregularizationfew-shotlearninggeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that negative prompt tuning improves out-of-distribution (OOD) detection on training classes but forgets pre-trained knowledge, which hurts detection on unseen classes and styles. The proposed method, Knowledge Regularized Negative Feature Tuning (KR-NFT), transforms pre-trained text features directly with light, image-conditional scaling and shifting parameters, and adds a knowledge-regularization loss that keeps the tuned features close to the pre-trained ones. Tuned on four-shot ImageNet samples, the method is claimed to reduce FPR95 by 5.44% against the closest competitor on unseen classes and by 16.77% on unseen styles, while also improving classification on the training distribution. The central question is whether OOD detection generalization can be achieved without sacrificing the pre-trained model's knowledge.

What carries the argument

The central object is the negative feature tuning transformation $\boldsymbol{c}'_i = L_2(\boldsymbol{\alpha} \boldsymbol{c}_i + \boldsymbol{\beta})$, applied separately to positive and negative text features, with $\boldsymbol{\alpha}$ and $\boldsymbol{\beta}$ generated per image by a lightweight meta-network from the image feature. The knowledge regularization loss $\mathcal{L}_{kr} = \frac{1}{N+M} \sum_{i=1}^{N+M} (1 - \boldsymbol{c}_i \cdot \boldsymbol{c}'_i)$ ties the tuned features to the pre-trained ones, preserving prior knowledge. At inference, the model uses the NegLabel score comparing similarity to the ID labels versus 10,000 mined negative labels, with the tuned text features replacing the pre-trained ones.

What would settle it

Take the same four-shot ImageNet-tuned KR-NFT model and re-run the unseen-class evaluation after removing from the 10,000 negative labels any word that matches a class name, synonym, or frequent object in CIFAR-10, CIFAR-100, Food-101, CUB, or Oxford Pets; if the FPR95 on unseen classes rises substantially toward the baseline, the generalization claim is an artifact of negative-label overlap rather than of the tuning itself.

Watch

Extended reading notes

Core claim

On the paper's own terms, KR-NFT establishes that a vision-language model can be adapted for OOD detection by directly modifying its frozen text features rather than by learning text prompts. The method applies element-wise scaling and shifting to the pre-trained text features, using separate transformations for positive (in-distribution) and negative (mined) text features, and generates the transformation parameters per image through a lightweight meta-network. A knowledge-regularization objective, which maximizes the cosine similarity between the pre-trained and the tuned text features, prevents the forgetting that the paper identifies in prior prompt-tuning methods. The paper demonstrates the approach by tuning on ImageNet with four samples per class and evaluating on unseen classes (CIFAR-10/100, Food-101, CUB, Oxford Pets) and unseen styles (ImageNet-R, ImageNet-A, ImageNet-V2, ImageNet-Sketch), reporting consistent FPR95 improvements in both settings.

Load-bearing premise

The reported gains on unseen classes rest on the assumption that the fixed 10,000 mined negative label words do not overlap with the names or typical contents of the unseen test classes; any such overlap would suppress in-distribution similarity scores and inflate the apparent detection improvement.

Editorial extensions

If this is right

  • Tuning text features directly, rather than optimizing prompt tokens, avoids backpropagation through the text encoder, cutting training TFLOPs to 0.11 versus 193.76 for conditional prompt tuning while keeping test speed fast.
  • The method is compatible with existing prompt-tuning methods; adding KR-NFT to LoCoOp, SCT, and LAPT improves their unseen-class and unseen-style FPR95.
  • OOD detection on unseen classes becomes a practical evaluation protocol: a model tuned on ImageNet can be applied to entirely new label spaces without retuning.
  • The knowledge-regularization weight $\lambda_2$ provides a single knob that trades base-class performance for generalization, with a moderate value balancing both.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same recipe—transform pre-trained features with small, conditional perturbations and penalize deviation from the original—could transfer to open-set recognition or continual learning, where catastrophic forgetting limits deployed vision-language models.
  • Because the meta-network reads only the image feature, one testable extension is to supervise the scaling and shifting factors on known shifted copies (e.g., blurred or sketch-like inputs) to see whether the conditional factors genuinely encode shift invariance.
  • A stronger evaluation would replace the single fixed 10,000-word negative vocabulary with a dynamic or class-aware one; if the reported gains persist, the method's robustness to negative-label choice would be established.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes Knowledge Regularized Negative Feature Tuning (KR-NFT) for out-of-distribution detection with vision-language models. Instead of learning prompt tokens, the method applies distribution-aware element-wise scaling and shifting directly to pre-trained text features, with image-conditional factors produced by a lightweight meta-network, and adds a knowledge regularization loss that keeps the transformed text features close to the pre-trained ones. The model is trained in a few-shot setting on ImageNet with positive and negative crops, and evaluated on standard OOD benchmarks as well as on unseen classes and unseen styles. The authors report consistent gains, including an average FPR95 of 22.79 on the ImageNet base setting, 34.99 on unseen classes, and 28.13 on unseen styles, together with compatibility experiments showing that KR-NFT can improve other prompt-tuning methods.

Significance. If the reported results hold, the paper addresses a genuine limitation of negative prompt tuning for OOD detection: such methods often improve in-distribution OOD detection while degrading on unseen classes and styles. The proposed feature-tuning formulation is simple, efficient, and well motivated, and the evaluation is unusually broad, covering standard OOD sets, unseen classes, unseen styles, the OpenOOD benchmark, multiple backbones, compatibility with other methods, and component ablations. The authors also release code and provide a computational-efficiency analysis. The main empirical claims, however, rest on the fixed NegLabel negative-label set and on a hyperparameter choice that uses the unseen-class evaluation itself, so the generalization evidence needs additional scrutiny before the headline numbers can be taken at face value.

major comments (3)
  1. [Section 3.1 / Section 4.1 / Table 2] The unseen-class generalization claim is evaluated with the NegLabel score of Eq. (2), whose denominator contains similarities to the fixed 10,000 negative labels Y−. The paper never checks whether Y− overlaps semantically with the label vocabularies of CIFAR-10/100, Food-101, CUB, or Oxford Pets. If any unseen class name or a near-synonym appears in Y−, genuine ID images of that class receive suppressed scores, and since each tuning method transforms negative features differently, the reported 5.44 percentage-point FPR95 advantage over LAPT on unseen classes could partly reflect differential handling of overlapping negative labels rather than genuine generalization. Please quantify the overlap (exact matches and top-k nearest negative labels for every unseen class), and re-run Tables 2 and 3 with any overlapping labels removed from Y− at inference, or otherwise demonstrate that the results are unchanged when Y− is guaranteed disjoint from all unseen label sets.
  2. [Section 4.3.1 / Fig. 4b] The hyperparameter λ2 is selected by jointly inspecting FPR95 on the Base (ImageNet) and New (CIFAR-10) settings, and the chosen value λ2=100 is then used for all results in Tables 1–3. Because CIFAR-10 is itself one of the datasets used to measure unseen-class generalization, the CIFAR-10 row of Table 2 is partially a test-set model-selection result. Please either select λ2 using only ImageNet validation data that is disjoint from the reported test evaluations, or provide full sensitivity curves for all unseen-class and unseen-style datasets to demonstrate that λ2=100 is not a peak chosen specifically on CIFAR-10.
  3. [Tables 1–3 / Fig. 4b] All results are reported as single numbers without error bars, significance tests, or the number of seeds. In the 4-shot training setup with stochastic crop selection (P=256, Q=32), run-to-run variation is likely non-negligible, and the headline improvements of 5.44% and 16.77% FPR95 are averages over heterogeneous datasets. Please report mean and standard deviation over at least three seeds for the main tables and the λ2 ablation, and state whether the main comparisons are significant, for example with paired tests.
minor comments (6)
  1. [Section 3.3 / Eq. (4)] The definition of β^neg appears to contain a typo: it reads {b^neg_1, ω^neg_2, ..., ω^neg_n} but should presumably be {b^neg_1, b^neg_2, ..., b^neg_n}.
  2. [Table 6] The column header 'FRR95↓' should read 'FPR95↓'.
  3. [Table 4] The citation for the row 'LAPT + KR-NFT' is given as [36], but LAPT is reference [81]; please correct the citation.
  4. [Figure 4a] The caption says 'hidden dimensions' while the panel and the surrounding text describe different backbones (RN50, ViT-B/16, ViT-L/14); please align the caption and axis labels with the actual experimental variable.
  5. [Section 2] The Related Work section contains the typo 'nvestigating' instead of 'investigating'.
  6. [Section 4.2] The text notes that results of other methods are referenced from their respective papers; please state the exact settings used for those references, including backbone, prompt template, and number of shots, to ensure comparability with the reported numbers.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: unseen-class/style results are measured on held-out datasets, and the self-cited LAPT baseline is not load-bearing evidence; the main caveats are hyperparameter selection on CIFAR10 and unexamined negative-label overlap, which are evaluation concerns rather than equation-level circular reductions.

full rationale

KR-NFT's text-feature transformation (Eqs. 4-6) and losses (Eqs. 7-10) are conventional training objectives: classification, OOD separation, and an explicit feature-consistency regularizer. The unseen-class and unseen-style numbers in Tables 2-3 are obtained by applying the trained model to held-out datasets, not by recovering the reported FPR95 values from fitted parameters or from the definitions of the loss terms. The knowledge-regularization loss (Eq. 9) transparently minimizes distance to pre-trained text features, so retaining zero-shot NegLabel-like behavior on unseen classes is an intended mechanism, not a hidden circularity. The closest baseline, LAPT [81], is prior work by overlapping authors, but it is used only as a comparison point and not as evidence for the method's premise. Two caveats are worth stating but do not amount to circularity: (1) lambda2=100 is selected by inspecting the Base/New tradeoff with CIFAR10 as the New set (Sec. 4.3.1, Fig. 4b), so the CIFAR10 row in Table 2 is partly a selected result; (2) the fixed 10,000 negative labels from NegLabel are not checked for overlap with unseen class or style vocabularies, which could affect the interpretation of Eq. (2) scores on those benchmarks. These are correctness and evaluation-validity concerns, not reductions of the derivation to its own inputs. Overall, the derivation chain is self-contained and externally benchmarked.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the reliability of CLIP's embedding space, on the fixed negative label set remaining disjoint from unseen ID classes, on the crop-selection heuristic separating foreground from background, and on several hyperparameters (lambda1, lambda2, Q, P, 10,000 negative labels) that are set without full sensitivity analysis. No new physical or conceptual entities are introduced; the meta-network is a standard neural network component.

free parameters (5)
  • lambda2 (knowledge regularization weight) = 100
    Chosen based on the Base/New H-MEAN tradeoff in Fig. 4b; higher values favor pre-trained knowledge retention at the cost of Base training performance.
  • lambda1 (OOD detection loss weight) = 0.3
    Fixed in all experiments; no sensitivity analysis is reported in the main text.
  • Q (number of selected positive and negative crops per sample) = 32
    Controls the size of the constructed positive and negative training sets; set without a reported ablation.
  • P (number of random crops per training sample) = 256
    Crop augmentation scale; fixed without a reported ablation.
  • Number of mined negative text labels = 10,000
    Taken from NegLabel; this value defines the negative label space used in both training and inference, and it is not varied in the paper.
assumptions (4)
  • domain assumption Pre-trained CLIP image and text features share a space where cosine similarity is a valid measure of semantic relevance.
    Relied on by the score function in Eq. (2), the crop selection in Section 3.4, and the knowledge regularization loss in Eq. (9).
  • domain assumption The 10,000 negative labels mined by NegLabel are a representative and semantically disjoint proxy for OOD data, including for unseen ID datasets used in the generalization evaluation.
    Adopted in Sections 3.1 and 4.1; the unseen-class FPR95 results depend on these labels not overlapping with the unseen class names.
  • domain assumption Crops with the highest cosine similarity to the class text are foreground/positive samples, and those with the lowest are background/negative samples.
    Used in Section 3.4 to construct D_p and D_n; if this heuristic fails for some classes, the training labels for L_p and L_n are noisy.
  • domain assumption The image-conditional meta-network outputs residual scaling and shifting that remain valid for classes and styles never seen during training.
    Section 3.3, Eq. (5); the generalization evaluation assumes the meta-network does not overfit the ImageNet training distribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Knowledge Regularized Negative Feature Tuning of Vision-Language Models for Out-of-Distribution Detection." pith.science (2026). https://pith.science/paper/RMHGM2U2

@misc{pith2026250719847,
  author       = {Pith},
  title        = {Pith review of: Knowledge Regularized Negative Feature Tuning of Vision-Language Models for Out-of-Distribution Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RMHGM2U2}},
  note         = {Machine review of arXiv:2507.19847}
}
read the original abstract

Out-of-distribution (OOD) detection is crucial for building reliable machine learning models. Although negative prompt tuning has enhanced the OOD detection capabilities of vision-language models, these tuned models often suffer from reduced generalization performance on unseen classes and styles. To address this challenge, we propose a novel method called Knowledge Regularized Negative Feature Tuning (KR-NFT), which integrates an innovative adaptation architecture termed Negative Feature Tuning (NFT) and a corresponding knowledge-regularization (KR) optimization strategy. Specifically, NFT applies distribution-aware transformations to pre-trained text features, effectively separating positive and negative features into distinct spaces. This separation maximizes the distinction between in-distribution (ID) and OOD images. Additionally, we introduce image-conditional learnable factors through a lightweight meta-network, enabling dynamic adaptation to individual images and mitigating sensitivity to class and style shifts. Compared to traditional negative prompt tuning, NFT demonstrates superior efficiency and scalability. To optimize this adaptation architecture, the KR optimization strategy is designed to enhance the discrimination between ID and OOD sets while mitigating pre-trained knowledge forgetting. This enhances OOD detection performance on trained ID classes while simultaneously improving OOD detection on unseen ID datasets. Notably, when trained with few-shot samples from ImageNet dataset, KR-NFT not only improves ID classification accuracy and OOD detection but also significantly reduces the FPR95 by 5.44\% under an unexplored generalization setting with unseen ID categories. Codes can be found at \href{https://github.com/ZhuWenjie98/KRNFT}.

Figures

Figures reproduced from arXiv: 2507.19847 by the authors.

Figure 1
Figure 1. GradCAM visualization of different methods on ID images. (a) In CLIP, the ID class shows high activation for both [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of our KR-NFT, where we introduce a knowledge regularized negative feature tuning with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of (a) Negative Prompt Tuning and (b) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Analyses on (a) hidden dimensions, (b) values of [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

92 extracted references · 45 canonical work pages

  1. [1]

    Davide Abati, Angelo Porrello, Simone Calderara, and Rita Cucchiara. 2019. Latent space autoregression for novelty detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 481–490

  2. [2]

    Syed Muhammad Anwar, Muhammad Majid, Adnan Qayyum, Muhammad Awais, Majdi Alnowami, and Muhammad Khurram Khan. 2018. Medical image analysis using convolutional neural networks: a review. Journal of medical systems 42 (2018), 1–13

  3. [3]

    Yichen Bai, Zongbo Han, Bing Cao, Xiaoheng Jiang, Qinghua Hu, and Changqing Zhang. 2024. ID-like Prompt Learning for Few-Shot Out-of-Distribution Detec- tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 17480–17489

  4. [4]

    Daniel Bogdoll, Maximilian Nitsche, and J Marius Zöllner. 2022. Anomaly detec- tion in autonomous driving: A survey. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 4488–4499

  5. [5]

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. 2014. Food-101–mining discriminative components with random forests. In Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13. Springer, 446–461

  6. [6]

    Chentao Cao, Zhun Zhong, Zhanke Zhou, Yang Liu, Tongliang Liu, and Bo Han. 2024. Envisioning Outlier Exposure by Large Language Models for Out-of- Distribution Detection. arXiv preprint arXiv:2406.00806 (2024)

  7. [7]

    Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and An- drea Vedaldi. 2014. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition . 3606–3613

  8. [8]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition . Ieee, 248–255

Show all 92 references
  1. [9]

    Li Deng. 2012. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine 29, 6 (2012), 141–142

  2. [10]

    Mohammad Mahdi Derakhshani, Enrique Sanchez, Adrian Bulat, Victor G Turrisi da Costa, Cees GM Snoek, Georgios Tzimiropoulos, and Brais Martinez. 2023. Bayesian prompt learning for image-language model generalization. In Proceed- ings of the IEEE/CVF International Conference on...

  3. [11]

    Terrance DeVries and Graham W Taylor. 2018. Learning confidence for out- of-distribution detection in neural networks. arXiv preprint arXiv:1802.04865 (2018)

  4. [12]

    Andrija Djurisic, Nebojsa Bozanic, Arjun Ashok, and Rosanne Liu. 2022. Ex- tremely simple activation shaping for out-of-distribution detection.arXiv preprint arXiv:2209.09858 (2022)

  5. [13]

    Xin Dong, Junfeng Guo, Ang Li, Wei-Te Ting, Cong Liu, and HT Kung. 2022. Neural mean discrepancy for efficient out-of-distribution detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 19217– 19227

  6. [14]

    Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. 2022. Unknown-aware object detection: Learning what you don’t know from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 13678–13688

  7. [15]

    Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. 2022. Zero-shot out-of-distribution detection based on the pre-trained model clip. In Proceedings of the AAAI conference on artificial intelligence , Vol. 36. 6568–6576

  8. [16]

    Stanislav Fort, Jie Ren, and Balaji Lakshminarayanan. 2021. Exploring the limits of out-of-distribution detection. Advances in Neural Information Processing Systems 34 (2021), 7068–7081

  9. [17]

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. 2024. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision 132, 2 (2024), 581–595

  10. [18]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

  11. [19]

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. 2021. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF internati...

  12. [20]

    Dan Hendrycks and Kevin Gimpel. 2016. A baseline for detecting misclassified and out-of-distribution examples in neural networks. arXiv preprint arXiv:1610.02136 (2016)

  13. [21]

    Dan Hendrycks, Xiaoyuan Liu, Eric Wallace, Adam Dziedzic, Rishabh Krishnan, and Dawn Song. 2020. Pretrained transformers improve out-of-distribution robustness. arXiv preprint arXiv:2004.06100 (2020)

  14. [22]

    Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. 2018. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606 (2018)

  15. [23]

    Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. 2019. Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781 (2019)

  16. [24]

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song

  17. [25]

    Geoffrey Hinton. 2015. Distilling the Knowledge in a Neural Network. arXiv preprint arXiv:1503.02531 (2015)

  18. [26]

    Rui Huang, Andrew Geng, and Yixuan Li. 2021. On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems 34 (2021), 677–689

  19. [27]

    Rui Huang and Yixuan Li. 2021. Mos: Towards scaling out-of-distribution de- tection for large semantic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 8710–8719

  20. [28]

    Dihong Jiang, Sun Sun, and Yaoliang Yu. 2021. Revisiting flow generative models for out-of-distribution detection. In International Conference on Learning Repre- sentations

  21. [29]

    Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, and Bo Han. 2023. Detecting out-of-distribution data through in-distribution class prior. In International Conference on Machine Learning . PMLR, 15067–15088

  22. [30]

    Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, and Bo Han. 2024. Negative label guided ood detection with pretrained vision-language models. arXiv preprint arXiv:2403.20078 (2024)

  23. [31]

    Di Jin, Shuyang Gao, Seokhwan Kim, Yang Liu, and Dilek Hakkani-Tür. 2022. Towards textual out-of-domain detection without in-domain labels. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30 (2022), 1386–1395

  24. [32]

    Diederik P Kingma. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

  25. [33]

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 2013. 3d object repre- sentations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops . 554–561

  26. [34]

    Alex Krizhevsky, Geoffrey Hinton, et al. 2009. Learning multiple layers of features from tiny images. (2009)

  27. [35]

    Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. 2018. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems 31 (2018)

  28. [36]

    Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. 2024. Learning transferable negative prompts for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 17584– 17594

  29. [37]

    Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. 2022. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems 35 (2022), 109–123

  30. [38]

    Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. 2017. Enhancing the reliabil- ity of out-of-distribution image detection in neural networks. arXiv preprint arXiv:1706.02690 (2017)

  31. [39]

    Ziqian Lin, Sreya Dutta Roy, and Yixuan Li. 2021. Mood: Multi-level out-of- distribution detection. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition . 15313–15323

  32. [40]

    Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. 2020. Energy-based out-of-distribution detection. Advances in neural information processing systems 33 (2020), 21464–21475

  33. [41]

    Xixi Liu, Yaroslava Lochman, and Christopher Zach. 2023. Gen: Pushing the limits of softmax-based out-of-distribution detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 23946–23955

  34. [42]

    Yifei Ming, Ziyang Cai, Jiuxiang Gu, Yiyou Sun, Wei Li, and Yixuan Li. 2022. Delving into out-of-distribution detection with vision-language representations. Advances in neural information processing systems 35 (2022), 35087–35102

  35. [43]

    Yifei Ming, Ying Fan, and Yixuan Li. 2022. Poem: Out-of-distribution detection with posterior sampling. In International Conference on Machine Learning . PMLR, 15650–15665

  36. [44]

    Yifei Ming, Yiyou Sun, Ousmane Dia, and Yixuan Li. 2022. How to exploit hyperspherical embeddings for out-of-distribution detection? arXiv preprint arXiv:2203.04450 (2022)

  37. [45]

    Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. 2024. Locoop: Few-shot out-of-distribution detection via prompt learning.Advances in Neural Information Processing Systems 36 (2024)

  38. [46]

    Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, An- drew Y Ng, et al. 2011. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning , Vol. 2011. Granada, 4

  39. [47]

    Anh Nguyen, Jason Yosinski, and Jeff Clune. 2015. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In Proceedings of the IEEE conference on computer vision and pattern recognition . 427–436

  40. [48]

    Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian

  41. [49]

    Aristotelis-Angelos Papadopoulos, Mohammad Reza Rajati, Nazim Shaikh, and Jiamian Wang. 2021. Outlier exposure with confidence control for out-of- distribution detection. Neurocomputing 441 (2021), 138–150. MM ’25, October 27–31, 2025, Dublin, Ireland Wenjie Zhu, Yabin Zhang, ...

  42. [50]

    Jaewoo Park, Yoon Gyo Jung, and Andrew Beng Jin Teoh. 2023. Nearest neigh- bor guidance for out-of-distribution detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 1686–1695

  43. [51]

    Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. 2012. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition . IEEE, 3498–3505

  44. [52]

    Stanislav Pidhorskyi, Ranya Almohsen, and Gianfranco Doretto. 2018. Generative probabilistic novelty detection with adversarial autoencoders. Advances in neural information processing systems 31 (2018)

  45. [53]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

  46. [54]

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. 2019. Do imagenet classifiers generalize to imagenet?. In International conference on machine learning. PMLR, 5389–5400

  47. [55]

    Jie Ren, Stanislav Fort, Jeremiah Liu, Abhijit Guha Roy, Shreyas Padhy, and Balaji Lakshminarayanan. 2021. A simple fix to mahalanobis distance for improving near-ood detection. arXiv preprint arXiv:2106.09022 (2021)

  48. [56]

    Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. 2019. Likelihood ratios for out- of-distribution detection. Advances in neural information processing systems 32 (2019)

  49. [57]

    Chandramouli Shama Sastry and Sageev Oore. 2020. Detecting out-of-distribution examples with gram matrices. In International Conference on Machine Learning . PMLR, 8491–8501

  50. [58]

    Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. 2012. Toward open set recognition. IEEE transactions on pattern analysis and machine intelligence 35, 7 (2012), 1757–1772

  51. [59]

    Yiyou Sun, Chuan Guo, and Yixuan Li. 2021. React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems 34 (2021), 144–157

  52. [60]

    Yiyou Sun and Yixuan Li. 2022. Dice: Leveraging sparsification for out-of- distribution detection. In European Conference on Computer Vision . Springer, 691–708

  53. [61]

    Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. 2022. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning. PMLR, 20827–20840

  54. [62]

    Leitian Tao, Xuefeng Du, Xiaojin Zhu, and Yixuan Li. 2023. Non-parametric outlier synthesis. arXiv preprint arXiv:2303.02966 (2023)

  55. [63]

    Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. 2019. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in neural information processing systems 32 (2019)

  56. [64]

    Joost Van Amersfoort, Lewis Smith, Yee Whye Teh, and Yarin Gal. 2020. Uncer- tainty estimation using a single deep deterministic neural network. In Interna- tional conference on machine learning . PMLR, 9690–9700

  57. [65]

    Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. 2018. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition . ...

  58. [66]

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie

  59. [67]

    Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. 2019. Learning Robust Global Representations by Penalizing Local Predictive Power. InAdvances in Neural Information Processing Systems . 10506–10518

  60. [68]

    Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. 2023. Clipn for zero-shot ood detection: Teaching clip to say no. InProceedings of the IEEE/CVF International Conference on Computer Vision . 1802–1812

  61. [69]

    Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. 2022. Vim: Out-of- distribution with virtual-logit matching. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 4921–4930

  62. [70]

    Hongxin Wei, Renchunzi Xie, Hao Cheng, Lei Feng, Bo An, and Yixuan Li. 2022. Mitigating neural network overconfidence with logit normalization. In Interna- tional conference on machine learning . PMLR, 23631–23644

  63. [71]

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

  64. [72]

    Jingkang Yang, Pengyun Wang, Dejian Zou, Zitang Zhou, Kunyuan Ding, Wenx- uan Peng, Haoqi Wang, Guangyao Chen, Bo Li, Yiyou Sun, et al. 2022. Openood: Benchmarking generalized out-of-distribution detection. Advances in Neural Information Processing Systems 35 (2022), 32598–32611

  65. [73]

    Geng Yu, Jianing Zhu, Jiangchao Yao, and Bo Han. 2024. Self-Calibrated Tuning of Vision-Language Models for Out-of-Distribution Detection. Advances in Neural Information Processing Systems 37 (2024), 56322–56348

  66. [74]

    Qing Yu and Kiyoharu Aizawa. 2019. Unsupervised out-of-distribution detection by maximum classifier discrepancy. In Proceedings of the IEEE/CVF international conference on computer vision . 9518–9526

  67. [75]

    Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. 2023. Task residual for tuning vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 10899–10909

  68. [76]

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. 2019. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision. 6023–6032

  69. [77]

    Alireza Zaeemzadeh, Niccolo Bisagno, Zeno Sambugaro, Nicola Conci, Nazanin Rahnavard, and Mubarak Shah. 2021. Out-of-distribution detection using union of 1-dimensional subspaces. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition . 9452–9461

  70. [78]

    Maxime Zanella and Ismail Ben Ayed. 2024. Low-rank few-shot adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 1593–1603

  71. [79]

    Jinsong Zhang, Qiang Fu, Xu Chen, Lun Du, Zelin Li, Gang Wang, Shi Han, Dong- mei Zhang, et al. 2022. Out-of-distribution detection based on in-distribution data patterns memorization with modern hopfield energy. InThe Eleventh International Conference on Learning Representations

  72. [80]

    Yabin Zhang and Lei Zhang. 2024. Adaneg: Adaptive negative proxy guided ood detection with vision-language models. Advances in Neural Information Processing Systems 37 (2024), 38744–38768

  73. [81]

    Yabin Zhang, Wenjie Zhu, Chenhang He, and Lei Zhang. 2024. Lapt: Label-driven automated prompt tuning for ood detection with vision-language models. arXiv preprint arXiv:2407.08966 (2024)

  74. [82]

    Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang

  75. [83]

    Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba

  76. [84]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 16816–16825

  77. [85]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Learning to prompt for vision-language models. International Journal of Computer Vision 130, 9 (2022), 2337–2348

  78. [86]

    Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Daeki Cho, and Haifeng Chen. 2018. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In International conference on learning represen- tations

  79. [87]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Dual memory networks: A versatile adaptation approach for vision- language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 28718–28728

  80. [2010]

    In 2010 IEEE computer society conference on computer vision and pattern recognition

    Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition . IEEE, 3485–3492

  81. [2011]

    The caltech-ucsd birds-200-2011 dataset. (2011)

  82. [2017]

    Places: A 10 million image database for scene recognition.IEEE transactions on pattern analysis and machine intelligence 40, 6 (2017), 1452–1464

  83. [2021]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Natural adversarial examples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 15262–15271

  84. [2024]

    In The Twelfth International Conference on Learning Representations

    Out-of-Distribution Detection with Negative Prompts. In The Twelfth International Conference on Learning Representations

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.