acceptodds
Withdrawn from ICLR 2027

The Confidence Illusion: Rethinking Entropy Minimization for Continual Test-Time Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continual Test-Time Adaptation (CTTA) aims to adapt a source model to continuously changing target domains using unlabeled test streams. Existing CTTA methods largely rely on entropy minimization, assuming that confident predictions provide reliable adaptation signals. However, we reveal a hidden confidence illusion: confident predictions are not always trustworthy. Under distribution shifts, misclassified samples can also exhibit low entropy and be repeatedly optimized, causing error accumulation. Moreover, such misleading confidence can gradually bias prediction distributions toward a few dominant classes, leading to long-term adaptation instability. To address this issue, we rethink entropy minimization from the perspective of reliable confidence optimization and propose Reliability-Aware Balanced Entropy (RBE). RBE distinguishes trustworthy confidence from misleading confidence through two complementary designs. Specifically, Consistency-aware Entropy Exploration (CAE) estimates prediction reliability by combining weak-augmentation consistency and teacher-student agreement, and selectively emphasizes reliable samples during entropy minimization. Meanwhile, Balanced Prediction Regularization (BPR) softly penalizes persistent concentration in a temporally aggregated prediction distribution, thereby mitigating class-dominant collapse. Extensive experiments on multiple CTTA benchmarks demonstrate that RBE achieves competitive overall performance and stable adaptation under continuously evolving target domains. Our findings highlight the importance of moving beyond confidence maximization toward reliable confidence modeling for continual adaptation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
continual test-time adaptation, domain adaptation, entropy minimization

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_QJZzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper tries to handle two problem in CTTA: confident predictions are not always trustworthy and bias prediction distributions toward a few dominant classes. To solve this, the authors propose Reliability-Aware Balanced Entropy and Consistency-aware Entropy Exploration. Experiments show that the method achieves competitive overall performance and stable adaptation under continuously evolving target domains.

Critical strengths
  1. The presentation of the paper is generally clear and easy to follow.
Critical weaknesses
  1. The biggest issue with this paper is its lack of novelty. Its motivation and the related techniques are already very common in the TTA literature. The reliability of low-entropy samples [1] and the class-collapse problem [2] have already been extensively studied, while consistency based on image augmentation [3] and teacher–student models [4] is also widely used in TTA. Overall, this paper brings me little new information.
  2. The experimental evaluation is insufficient in scope. The authors do not evaluate the proposed method under several important CTTA settings [5]. I suggest that the authors supplement the paper with experiments under these relevant settings.

[1] Entropy is not Enough for Test-Time Adaptation: From the Perspective of Disentangled Factors. ICLR 2024

[2] NOTE: Robust Continual Test-time Adaptation Against Temporal Correlation. NeurIPS 2022

[3] MEMO: Test Time Robustness via Adaptation and Augmentation. NeurIPS 2022

[4] Continual Test-Time Domain Adaptation. CVPR 2022

[5] Towards Stable Test-time Adaptation in Dynamic Wild World. ICLR 2023

Questions for authors

see weaknesses

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

The paper proposes Reliability-aware Balanced Entropy (RBE) to address confirmation bias in Continual Test-Time Adaptation (CTTA) by revisiting entropy minimization. RBE estimates reliability by combining augmentation consistency and model consistency. In addition, the authors introduce Balanced Prediction Regularization (BPR), which discourages persistent concentration in the temporally aggregated prediction distribution to mitigate class collapse in CTTA. The proposed method achieves improved performance across various corruption datasets.

Critical strengths
  • The paper has a clear motivation to improve entropy-based reliable sample selection strategies and achieves strong performance across various experiments.
  • The paper tried to revisit the assumption of entropy minimization in continual test-time adaptation task, and the consistency-aware entropy exploration that transforms entropy minimization from confidence-driven optimization into reliability-guided optimization by continuously estimating sample-wise optimization reliability.
  • The paper proposed the balanced prediction regularization that penalizes persistent concentration in a temporally aggregated prediction distribution.
Critical weaknesses
  • From the viewpoint of concerns for Novelty. Augmentation-based entropy minimization strategies have already been explored in previous methods, such as DeYO [1] and SPA [2], which achieve efficiency without requiring a teacher model. Prediction bias correction based on the class distribution has also been studied in PTA [3]. Furthermore, the paper lacks comparisons with methods that address the limitations of entropy-based sample selection strategies, such as ReCAP [4]. The contribution could be thought as a simple combination of the previous works.
  • Another concern on Evaluation fairness is that the training regimes differ between the proposed method and the baselines, making it difficult to ensure a fair comparison. In particular, the selection of trainable parameters and detailed learning rate tuning may contribute to the performance improvements independently of the proposed method. Without solving this concerns, nobody think the proposed method works successfully in CTTA tasks.
  • In applicationability, existing CTTA methods, such as CoTTA, ViDA, and GOLD, report results on semantic segmentation, which is important for practical CTTA applications such as autonomous driving. However, it remains unclear whether the proposed method is also effective for semantic segmentation.

[1] Entropy is not Enough for Test-Time Adaptation: From the Perspective of Disentangled Factors, ICLR 2024.

[2] Self-Bootstrapping for Versatile Test-Time Adaptation, ICML 2025.

[3] Partition-Then-Adapt: Combating Prediction Bias for Reliable Multi-Modal Test-Time Adaptation, NeurIPS 2025.

[4] Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation, ICML 2025.

Questions for authors
  1. How does the number of augmentations affect the performance of the proposed method?

  2. It is unclear why the proposed method is considered memory-efficient compared to parameter-efficient fine-tuning methods such as ViDA and DPCore.

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

This paper investigates the reliability of entropy minimization in Continual Test-Time Adaptation. The authors point out that continuously optimizing confidently incorrect predictions under evolving target domains may lead to error accumulation and degeneration of the predicted class distribution. To address this issue, the paper proposes Reliability-Aware Balanced Entropy (RBE), which consists of two main components: Consistency-aware Entropy Exploration (CAE) and Balanced Prediction Regularization (BPR). Experiments are conducted on three datasets with two backbone architectures, demonstrating generally promising results.

Critical strengths
  1. The paper has a clear motivation and addresses a problem with practical significance.

  2. The proposed method is easy to understand.

  3. The paper provides extensive experiments and achieves generally promising results.

Critical weaknesses
  1. Ranked Entropy Minimization (REM) has already investigated entropy-induced instability and model collapse in CTTA, which substantially overlaps with the motivation of this paper. Moreover, BPR combines existing techniques and is closely related to conventional information maximization and prediction-diversity regularization. These similarities raise concerns about the distinctiveness and novelty of the proposed method.

  2. Despite the extensive experiments, the paper lacks comparisons with some of the most closely related baselines.

Questions for authors
  1. REM [1] and the proposed RBE share highly similar motivations. Why is REM not discussed in this paper?

  2. RBE combines several existing techniques. What makes this combination a meaningful methodological advancement over prior work?

  3. The paper should include comparisons with closely related methods, particularly the recent REM, DeYO [2] and SaTeen [3]. Furthermore, although the related work section is relatively extensive, it lacks sufficient discussion of these highly relevant approaches.

  4. The proposed method performs relatively poorly on CIFAR10-C. Could the authors provide further analysis to explain this performance gap?

  5. I suggest using vector graphics for the experimental figures to improve their visual quality and readability.

[1] Ranked entropy minimization for continual test-time adaptation. ICML 2025.

[2] Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. ICLR 2024.

[3] SaTeen: Learning Structural Alignment for Continual Test-Time Adaptation. ICML 2026.

Reviewer_jk6POpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies error accumulation in continual test-time adaptation (CTTA), focusing on the observation that low-entropy predictions can be incorrect and that minimizing their entropy may reinforce these errors. It proposes Reliability-Aware Balanced Entropy (RBE), which combines two components. Consistency-aware Entropy Exploration (CAE) weights each sample’s entropy loss using consistency under weak augmentation and agreement between the student and an EMA teacher. Balanced Prediction Regularization (BPR) maintains an exponential moving average of the predicted class distribution and penalizes its KL divergence from a uniform distribution to discourage persistent concentration on a few classes.

Critical strengths
  1. Distinguishing confidence from reliability is well motivated. CAE uses two consistency signals rather than treating confident predictions as inherently trustworthy, and the resulting weighting scheme is straightforward to integrate into entropy-based adaptation.

  2. RBE performs well on ImageNet-C and CIFAR100-C and is evaluated with both CNN and Transformer backbones. The paper also reports its weaker result relative to DAS on CIFAR10-C rather than claiming uniformly superior performance.

Critical weaknesses
  1. Consistency weighting, EMA teachers, and marginal prediction regularization are well-established ideas. Comparisons against simple confidence-based weighting and batch-level marginal entropy maximization are needed to establish the value of the specific CAE and BPR designs.

  2. Regardless of whether uniformity is described as a reference or a target, minimizing Equation (11) pushes the temporally aggregated predictions toward a uniform distribution. This may conflict with correct predictions under sustained class imbalance or label shift, which the current benchmarks do not adequately test.

  3. Figure 3 evaluates reliability on Gaussian noise without comparing against entropy-based ranking. Lower error among low-entropy predictions in Figures 1 and 4 could also follow from improved overall accuracy, rather than specifically demonstrating suppression of confidently incorrect updates.

  4. The paper does not report variation across seeds or corruption orders and leaves important details unclear, including weak augmentations, teacher momentum, trainable parameters, and hyperparameter selection. These omissions matter particularly for the 0.8-point improvement over the strongest ResNet-50 baseline.

Questions for authors
  1. Under matched settings, how does CAE compare with continuous weighting based on entropy or maximum predicted probability? How does BPR compare with marginal entropy maximization without temporal averaging? These controls would clarify which parts of the proposed design are responsible for the gains.

  2. Could you provide results on class-correlated streams or persistent label-distribution shifts? Temporal averaging can reduce short-term batch fluctuations, but it does not resolve a mismatch between the uniform reference and a genuinely nonuniform target distribution.

  3. Within the low-entropy subset, how do weights differ between correct and incorrect predictions? Correctness-ranking AUROC across multiple corruption domains, with entropy as a baseline, would be informative. Stable errors shared by the student and EMA teacher are particularly relevant to the paper’s central claim.

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

Related papers

Discussion (0)

Sign in to comment.