acceptodds
Under review as a conference paper at ICLR 2027

RELIABILITY-GATED TEST-TIME ADAPTATION FOR OPEN-WORLD VISION-LANGUAGE RECOGNITION

Abstract

Pretrained vision-language models provide strong zero-shot recognition but remain brittle when deployment streams combine evolving domain shifts with categories absent from the task label set. Existing test-time adaptation methods typically update on confident predictions in a closed label space, making open-world streams especially vulnerable to confirmation bias and irreversible drift. We present REG TTA, a reliability-gated framework for online adaptation of frozen vision-language models. REG-TTA scores each incoming sample using prediction consistency across stochastic augmentations, image-text semantic agreement, and local feature density in a dual-timescale prototype memory. Only reliable known-class instances update a lightweight prompt adapter, while uncertain or potentially unknown inputs are excluded from entropy minimization and organized into temporary semantic prototypes. We analyze the gradient induced by unknown-class samples and show that closed-set entropy minimization introduces a systematic pseudo class bias whenever the model is overconfident on open-set inputs. The released supplementary package contains an implementation,public-dataset stream builders, aggregation scripts, and plotting utilities for ImageNet-C, ImageNet-R, ImageNet A, ImageNet-V2, ObjectNet, and DomainNet evaluation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Test‑Time Adaptation, Vision‑Language Models, Open‑World Recognition, Prompt Tuning, Out‑of‑Distribution Detection

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 4 reviews, above 0% of the others.
Reviewer_tL3eOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces a test-time adaptation (TTA) method for vision-language models (VLMs) under distribution shifts in an open-world setting, i.e., samples from unknown classes may be encountered during testing. The paper proposes a reliability gate and a dual-timescale prototype memory to select known and confident samples for parameter updates and to maintain persistent class prototypes, respectively. However, this paper is incomplete because it does not provide experimental results.

Critical strengths
  • S1: The problem setting where distribution shifts and unknown classes simultaneously occur is realistic.
Critical weaknesses
  • W1: This paper is incomplete and not yet ready to be reviewed:
    • Experimental results are not provided.
    • The layout of figures is collapsed.
    • The AI use statement is not included, which violates the guideline.
Questions for authors

N/A

Reviewer_B2ceOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

I recommend that the AC escalate this under the ICLR 2027 AI policy. Please block this account.

Critical strengths

rubbish paper.

Critical weaknesses

My assessment: this paper was generated by a large language model and has not been verified by any human. There are no experimental results in this paper. I read all 10 pages and scanned the full text. The sections are 1 Intro → 2 Related Work → 3 Problem Setup → 4 Method → 5 Evaluation Protocol → 6 Discussion → 7 Conclusion. There is no Experiments/Results section at all.

Questions for authors

Please leave the academic community immediately and do not waste the reviewers' time !!!!!!!!!!!!!

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

This paper proposes a method to improve test-time adaptation for vision-language models in open-world settings, where only realiable instances are used to update a lightweight adapter to the frozen, reference, VLM. Reliability is determined using the dot product between normalised image and text features. There are no results presented.

Critical strengths

The problem discussed in the paper is important and motivated.

Critical weaknesses

1/ There are serious problems with the presentation of the paper. In particular:

  • Figure 1: Text written over other text
  • Figure 2: Text written over other text. Also, it is entirely unclear what the lines and dots in the figures are supposed to represent. I find it hard to believe that this was re-read by someone, and feel that the figures 1-2 were not manually verified.

2/ Even more problematic, there are no results ! There is only a description of an experimental protocol, no actual results. The same goes for the ablations, there are no quantitative evaluations. Again, I have a hard time believing this was re-read.

Other comments:

  • equation 1: τ\tau not defined;
  • line 120: what are the stochastic augmentations?
Questions for authors

1/ My main comment would be for the authors to re-read the paper, it is, at the moment, impossible to judge and difficult to understand, in large part due to the poorly realised diagrams and the lack of any results.

2/ Why is the dot product between normalised image and text features used as reliability? This is never explained clearly.

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

The paper proposes REG-TTA for vision-language models facing both distribution shifts and unknown classes at test time. It uses prediction consistency across augmented views, image–text agreement, and prototype density to decide which samples can safely update a prompt adapter. A long-term and a short-term memory track class information and recent changes in the stream.

Critical strengths

The paper addresses a relevant failure mode: updating a model on unknown-class samples may cause it to drift toward incorrect known-class predictions. Its reliability signals and online evaluation metrics are described clearly enough to understand the proposed approach.

Critical weaknesses
  1. The main paper reports no numerical performance results. Table 1 lists datasets, Table 2 defines metrics, Table 3 describes the parameter budget, and Figure 4 shows the reporting pipeline. None reports known-class accuracy, unknown-class AUROC or FPR95, forgetting, runtime, or a component ablation. The evaluation plan alone cannot support a claim that REG-TTA is effective.
  2. The stated baseline set omits closely related methods. The main paper does not explain the absence of STAMP, UniEnt, or CLIPTTA. These works address open-set sample selection, unknown-class handling, or vision-language test-time adaptation. Without a comparison or an explanation of why they do not fit the protocol, it is difficult to assess the contribution of REG-TTA.
Questions for authors

Regarding weakness 1: Are numerical results available in a separate supplementary file? If so, where are the comparison tables and ablations, and why are the central results absent from the main paper? In particular, are known-class accuracy, unknown-class AUROC/FPR95, forgetting, runtime, and the contributions of the three reliability signals and dual-timescale memory available? Regarding weakness 2: What criteria were used to select the baselines? Which aspects of the proposed stream protocol prevent direct comparison with STAMP, UniEnt, or CLIPTTA? If comparable results are available, they would help establish the benefit of the proposed method.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.