acceptodds
Withdrawn from ICLR 2027

Online Semantic Gaussian Banks for Test-Time OOD Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pre-trained vision-language models (VLMs) have shown strong performance for out-of-distribution (OOD) detection, while recent test-time methods further improve detection by adapting to unlabeled test streams. However, incorrect ID/OOD assignments during adaptation can introduce unreliable test samples and degrade detection performance. We propose OSGB (Online Semantic Gaussian Banks), an optimization-free framework for online test-time OOD detection. Using frozen VLM features and positive and negative textual semantics, OSGB selectively updates separate Gaussian banks for ID and OOD samples. As test-time statistics accumulate, a global Gaussian likelihood-ratio score is used for OOD detection. Each sample is scored before updating the banks, preserving causal online inference. Experiments on ImageNet-1K, shifted ImageNet variants, CIFAR-100, and OpenOOD show consistent improvements over the compared methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Out-of-Distribution Detection, Test-Time Adaptation, Vision-Language Models, Online Adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_iiaAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies online test-time out-of-distribution (OOD) detection with frozen vision-language models. The authors focus on a central difficulty in adapting to unlabeled mixed ID/OOD streams: incorrect assignments can contaminate the adaptation state and reinforce subsequent detection errors. To address this issue, the paper proposes OSGB, an optimization-free framework that maintains separate ID and OOD Gaussian banks in frozen CLIP feature space. Positive class names and negative semantic concepts guide candidate selection, while several rules control bank updates. Experiments cover ImageNet-1K and other benchmarks, with strong reported performance.

Critical strengths
  • The paper addresses a practical challenge in online OOD detection: learning from an unlabeled stream without allowing unreliable assignments to corrupt future predictions. The proposed method keeps CLIP frozen and updates sufficient statistics to identify reliable samples instead of using gradient-based test-time adaptation.
  • OSGB achieves substantial improvements across several evaluated settings. On ImageNet-1K, it outperforms the AdaNeg and TTL baselines. The proposed method also performs well on other benchmarks.
  • The appendix includes global Gaussian baselines, bootstrap ablations, and an analysis of OOD-bank purity and coverage. Code is also provided to reproduce the results.
Critical weaknesses
  • The definition of adaptive-z3 is unclear. Does it refer to Equation 1?
  • The complete detector performs strongly, but the additional contribution of semantic Gaussian decomposition appears modest. In Table 7, replacing global Gaussian banks with semantic banks changes the average FPR95 from 6.42% to 6.11% and AUROC from 98.33% to 98.34%. Without uncertainty estimates, it is difficult to assess whether this improvement justifies the additional components.
  • The main ImageNet experiments use randomly ordered streams with a single fixed OOD source and predominantly ID samples. Achieving strong performance requires accumulating a substantial number of samples: the average FPR95 is 22.09% after 2K samples, compared with 6.11% over the full stream. This setting does not reflect real-world streams, in which different OOD sources may appear over time.
  • It is unclear why TTL’s performance in Table 1 is lower than that reported in its original paper, whereas its performance in Table 2 matches the original results.
  • Figure 1(a) is broken.
  • Table results is far from the experiment section.
Questions for authors
  1. How is adaptive-z3 defined, and does it correspond to Equation 1?
  2. Are the gains from semantic Gaussian decomposition statistically significant across multiple runs, and do they hold across other datasets?
  3. How does OSGB perform when OOD sources change over time or when only a small number of samples are available for adaptation?
  4. Why is TTL’s performance in Table 1 lower than that reported in its original paper, while its performance in Table 2 matches the original results? Were the evaluation settings or implementation details different?
Reviewer_1uFROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

To address incorrect ID/OOD assignments during adaptation can introduce unreliable test samples and degrade detection performance. this paper proposes OSGB (Online Semantic Gaussian Banks), an optimization-free framework for online test-time OOD detection.

Critical strengths
  1. addresing incorrect ID/OOD assignments during adaptation is important in test-time adaptation based OOD detection.

  2. This paper is well written and easy to follow

  3. Discussion on related work is informative.

Critical weaknesses
  1. In table 7, the authors shows that using one global Gaussian per bank results in marginal performance change (FPR95 from 6.11% to 6.42% and AUROC from 98.34% to 98.33%). Given that Online Gaussian modelling of CLIP features is already proposed in [1,2], this observations deteriorates the contribution of the proposed semantic decomposition.

  2. The reproduced TTL results also differ materially from the published results: average FPR95 changes from 12.46% to 15.52% on ImageNet and from 2.36% to 8.79% on CIFAR-100, which makes the experiment comparision unfair.

  3. the authors claim in section 4.4 that The exact timing convention, post-encoder overhead, and memory measurements are provided in the supplementary material. however, these details are absent from the supplied appendix.

  4. In Table 8, the authors show that RAF improves bank purity but leads to worse FPR95, which is in conflict with the claim "cleaner banks produce better detection".

  5. While I agree that the need of incorrect ID/OOD assignments during adaptation, there lacks theoretical analysis explaining under what conditions the proposed method reduce contamination instead of reinforcing errors introduced during early adaptation.

Besides, there lack a thereoctical analysis why the proposed method can addressing incorrect ID/OOD assignments during adaptation. As a result, it is not convinicng the effectivenss of the proposed method. As a result, it is unclear that these gains arise from addressing the assignment errors that motivate the work.

[1] Test-Time Retrieval-Augmented Adaptation for Vision-Language Models, ICCV 2025 [2] Online Gaussian Test-Time Adaptation of Vision-Language Models, CVPRw2025

Questions for authors

see weakness

Reviewer_CS2oOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OSGB, an optimization-free framework for online test-time OOD detection that uses frozen CLIP features and textual semantics to maintain separate ID and OOD Gaussian banks.

Critical strengths
  1. The framework requires no gradient-based optimization and preserves causal online inference by scoring each sample before updating its statistics.

  2. Experimental results shows taht OSGB achieves strong average performance across multiple standard and shifted OOD benchmarks.

  3. The paper is well written and easy to follow.

Critical weaknesses
  1. The central semantic decomposition offers only 0.31% FPR95 reduction and 0.01% AUROC gain over global Gaussian bank management, leaving its added complexity insufficiently justified.

  2. Robustness to evolving streams remains limited, with FPR95 reaching 23.86% under non-stationarity and 22.28% under an OOD-dominated stream, raising concerns about applicability to dynamic online settings.

  3. Tab. 4 does not specify the detection score used when global Gaussian scoring is disabled, although bank updates alone cannot change the frozen semantic score in Eq.(2), making the reported component-wise improvements difficult to interpret.

Questions for authors

Refer to the weakness section.

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

Related papers

Discussion (0)

Sign in to comment.