acceptodds
Withdrawn from ICLR 2027

CC-TTRL: Control-Conditioned Test-Time Selection of Normalization Statistics under Label Shift

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Test-time adaptation methods based on Batch Normalization can correct batch-specific technical variation using unlabeled target samples, but their statistics become biased when the target class distribution shifts. In biomedical imaging, negative controls provide a stable reference for separating technical variation from perturbation-specific signal. CS-ARM-BN exploits this structure by pooling controls and perturbed samples when recomputing BatchNorm statistics, yielding a favorable global bias-variance trade-off. However, the same estimator is applied to every normalization layer and every target batch, despite the fact that class-dependent activation bias varies across network depth and target composition. We introduce CC-TTRL, a control-conditioned test-time adaptation method that selects the source of normalization statistics independently across network stages. For each stage, CC-TTRL chooses among pooled target statistics, control-only statistics, and stored training statistics, producing a discrete family of adaptations that includes CS-ARM-BN as a special case. Candidate configurations are evaluated without target labels using a cross-fitted reward that combines agreement between held-out target controls and source controls with prediction consistency on held-out perturbed samples. A trust-region constraint and improvement veto preserve the CS-ARM-BN solution when evidence for an alternative configuration is insufficient. We search this action space using a lightweight policy-gradient procedure. We evaluate CC-TTRL on mechanism-of-action prediction from high-content cellular microscopy under experimental-batch shift and label shift. Under severe label shift, CC-TTRL improves over CS-ARM-BN and preferentially avoids pooled statistics in deeper network stages, consistent with increased class-dependent activation leakage at greater depth. The proposed control-anchored reward correlates with held-out predictive accuracy and substantially outperforms a feature-space reward. Random search with the same reward and evaluation budget achieves comparable performance to policy-gradient search, indicating that the primary contribution is the layerwise adaptation space and label-free control-guided selection criterion rather than reinforcement learning itself.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning, Test-Time Adaptation, Domain Adaptation, Distribution Shift, Biomedical Imaging, Adaptive Normalization, Mechanism-of-Action Prediction

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 3 reviews, above 0% of the others.
Reviewer_xncTOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors used reinforcement learning to improve upon CS-ARM-BN for test time adaptation in biomedicine by treating the layer norm parameters in different layers as different settings to be evaluated.

Critical strengths

The idea to use re-inforcement learning to improve upon test-time adaptation is to the best of my knowledge novel and interesting. Defining discrete sets to limit the search space for the re-inforcement learning is interesting and makes sense in the context of biomedicine.

Critical weaknesses

One critical weakness is the lack of care for the manuscipt:

  • The paper is unfinished as the authors were extremely wasteful with the usage of space (Line 291 is just p_src) 4 pages before the start of the method section. Still, the authors do not manage to even fill up the 9 pages and are missing even basic information such as which dataset they used to train/evaluate their model. They are also missing mandatory section such as the AI Use statement.
  • The paper is also missing citations (the entire related work section has multiple ?? instead of citations), which stem from errors in the latex.
  • The paper also lacks any description of their implementation details and examples of the data used and any quantitativ results.
  • Their recomputed values lack any standard deviation, while all other values have them.
  • Lack of comparison methods. They compare themselves with a single method on a single dataset and use the method as a fallback in case the performance did not increase.

Another critical weakness is the general assumption that using non-pooled statistics is better than pooled statistics. The authors did not show that there even is a configuration that outperforms just using pooled statistics everywhere. The final result is marginally better than just using CS-ARM-BN and no upper bound is provided.

Questions for authors

The manuscript would need to be completely rewritten for me to materially change my opinion, which is not the aim of the rebuttal.

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

This paper proposes CC-TTRL, a test-time adaptation method for classifying the mechanism of action of drugs from cellular microscopy images. The method aims to improve model performance when the distribution of classes changes between training and testing. Such changes can affect the Batch Normalization (BN) statistics used by the model and lead to less accurate predictions.

The proposed method builds on CS-ARM-BN, which calculates BN statistics using both untreated control samples and drug-treated samples. Instead of using the same approach throughout the network, CC-TTRL selects between three types of BN statistics at different network stages: statistics calculated from both controls and drug-treated samples, statistics calculated from controls only, and statistics stored during training. Since the true labels of the test samples are unavailable, the authors introduce a reward function based on the predictions of control samples and the agreement between predictions obtained using different BN configurations. They use a policy-gradient-based search to identify a suitable configuration without updating the model weights.

The authors evaluate their method under different experimental conditions and levels of label shift. The results show slight improvements over CS-ARM-BN mainly when the label distribution changes substantially. However, the method provides little or no improvement in settings with less severe shifts. The authors also report that random search performs similarly to their policy-gradient-based search.

Critical strengths
  • The paper addresses an interesting and relevant problem by investigating the impact of label shift on Batch Normalization-based test-time adaptation.

  • Selecting different BN statistics at different network stages is an intuitive and a reasonable extension of CS-ARM-BN.

  • The paper evaluates the proposed method under multiple experimental settings and compares its performance against CS-ARM-BN.

Critical weaknesses

1-Limited improvement over the baseline:

The main concern is that the proposed method provides only modest improvements over CS-ARM-BN. Given the additional complexity of the proposed method, the experimental results do not convincingly demonstrate a substantial advantage over the baseline.

2-Limited justification for using prediction agreement as a selection criterion:

The proposed reward assumes that a normalization configuration is more reliable when its predictions agree with the majority of evaluated configurations. However, agreement does not necessarily indicate correctness, particularly because all configurations use the same pretrained network and differ only in their BN statistics. They may therefore share systematic prediction errors. The authors should provide stronger empirical evidence that this agreement measure is a reliable indicator of classification accuracy.

3-Insufficient statistical analysis of the results:

Table 1 reports mean accuracy and standard deviation for CC-TTRL and the published CS-ARM-BN results, but only mean accuracy for the recomputed CS-ARM-BN baseline. Since the reported improvements are calculated relative to the recomputed baseline, its variability is particularly important. Without standard deviations or paired statistical significance tests, it is difficult to assess whether the reported improvements are reliable.

4-Limited contribution of the policy-gradient search:

The authors report that random search achieves essentially the same performance as the proposed policy-gradient search (0.909 versus 0.909, p=0.87p=0.87). This raises questions about the necessity of the proposed search strategy. If a simpler random search performs equally well, the authors should clarify what advantage the policy-gradient approach provides and whether its additional complexity is justified.

5-Limited evaluation across datasets:

Although the paper evaluates several experimental settings, the experiments are conducted within one cellular microscopy application using two data sources, S3 and S8. The different label shift settings are variations of the same underlying evaluation rather than independent datasets. Therefore, it is unclear whether the proposed method generalizes to other datasets.

Minor points

  • The Related Work section contains numerous unresolved citations, displayed as question marks.

  • Table 1 has some formatting issues.

Questions for authors

1- The improvements over CS-ARM-BN are relatively modest and are mainly observed under severe label shift. Could the authors elaborate on the practical significance of these improvements? Why selecting the batch norm at every layer does not lead a more significant improvement?

2- The proposed reward relies partly on agreement between predictions from different BN configurations. Since all configurations use the same pretrained model, they may share systematic prediction errors. Could the authors provide empirical evidence that prediction agreement is correlated with classification accuracy?

3- Why are standard deviations reported for the published CS-ARM-BN results and CC-TTRL, but not for the recomputed CS-ARM-BN baseline in Table 1? Could the authors provide variability estimates for the recomputed baseline and paired statistical significance tests or confidence intervals for the reported improvements?

4- The authors report that random search achieves comparable performance to the proposed policy-gradient search (0.909 versus 0.909, p=0.87). What is the advantage of using policy-gradient search over a simpler random search?

5- The evaluation is limited to one cellular microscopy application with two data sources (S3 and S8). Could the authors evaluate the proposed method on additional independent datasets to demonstrate its generalizability?

Reviewer_ruk6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper presents CC-TTRL, a reinforcement learning-based test-time adaptation method for biomedical imaging. The method selects normalization statistics independently across network stages without requiring target labels. Specifically, CC-TTRL chooses among pooled target statistics, control-only statistics, and stored training statistics. The selection is guided by a cross-fitted, label-free reward that combines the alignment of target negative-control predictions with source-control references and prediction agreement on perturbed samples. A lightweight group-normalized policy-gradient algorithm is used to explore the discrete space of normalization configurations.

Critical strengths

Unfortunately, I do not see strengths in this submission.

Critical weaknesses
  • Poor presentation of the paper: The paper suffers from significant presentation and writing issues that make it difficult to follow and understand. For instance, several citations are missing in the Related Work section, where placeholders (???) appear instead of proper references. Furthermore, some mathematical functions and notations are just introduced without adequate definitions or explanations. For example, the (JS) function used to measure alignment in line 296 is not defined. In addition, equations are not numbered, making it difficult to refer to them throughout the paper. Also, table 1 is poorly organized making it difficult to interpret.

  • Unjustified design choices: Several design choices are introduced without sufficient motivation or justification. For instance, in lines 154-155 authors claim that a random search strategy with the same evaluation budget reaches the same accuracy as the used reinforcement learning based policy. This raises questions about the necessity for using a more complex policy-gradient optimization procedure when a simpler random search strategy performs equally well. Similarly, the trust-region threshold of 0.2 and the improvement margin of 0.01 appear to be arbitrarily selected, without any sensitivity analysis to assess their impact on performance. Furthermore, eight configurations per round over three rounds is not adequately justified.

  • Missing ablation study of the reward function: The proposed reward function combines three components: source–target control alignment, consistency within target controls, and prediction agreement on perturbed samples. However, the paper does not provide any ablation study to assess the individual contribution of each component. Therefore, it is unclear whether all three components are necessary or whether similar performance could be achieved using a simpler reward formulation.

  • Missing implementation details and reproducibility concerns: The paper lacks sufficient implementation details to ensure reproducibility, such as the hyper-parameters used for policy-gradient optimization. Furthermore, no source code is provided, nor do the authors indicate whether they intend to release their implementation.

  • Concerns regarding the results in Table 1: First, the experimental settings presented in Table 1 are difficult to distinguish and interpret. Second, there are noticeable discrepancies between the published and recomputed CS-ARM-BN baseline results (e.g., 0.776 vs. 0.740 for the S8→S3 setting), without any explanation for these differences. Third, the reported improvements are relatively small in several cases, and no statistical analysis is provided to determine whether these improvements are statistically significant. Furthermore, the proposed method should be compared against a broader range of relevant test-time adaptation baselines, rather than only CS-ARM-BN, to provide a more comprehensive evaluation of it.

Questions for authors

I do not have further questions for the authors.

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

Related papers

Discussion (0)

Sign in to comment.