acceptodds
Under review as a conference paper at ICLR 2027

LAST: Stabilizing Test-Time Adaptation via Logit Standardization

Abstract

Test-time adaptation (TTA) improves model performance under distribution shift by leveraging unlabeled test data during inference. However, existing TTA approaches often suffer from model collapse, where predictions progressively concentrate on a limited subset of classes. Despite its prevalence, the underlying mechanism of this failure mode remains insufficiently understood. In this work, we investigate the dynamics of TTA collapse and identify logit norm instability as a key factor driving prediction degeneration. We show that adaptation-induced logit norm drift disturbs the balance of prediction confidence across classes, amplifying inherent prediction biases through a self-supervised process. Specifically, excessively increased logit norms lead to overconfident predictions that strengthen erroneous class preferences, while reduced logit norms weaken class separability and impair stable adaptation. Based on this insight, we propose **L**ogit **A**daptation via **S**tandardization at **T**est time (**LAST**), a lightweight and plug-and-play module that performs class-wise logit standardization before the layer. By normalizing the distribution of logits for each class across test samples, LAST suppresses instance-dependent logit norm fluctuations and regulates prediction confidence. Extensive experiments across disserve benchmarks show that LAST consistently improves existing TTA methods and has ability to alleviate collapse, achieving up to a 36.5% improvement in Top-1 accuracy over TENT baselines.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Test-time adaptation, Distribution shift, Model prediction collapse, Logit dynamics, Confidence calibration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Q5M7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper aims to address the tendency of predictions to become increasingly concentrated on a small subset of classes during TTA by estimating the mean and standard deviation of each class’s logits and standardizing them accordingly.

Critical strengths

The method is simple and easy to understand. By standardizing class-wise logit distributions, it may help mitigate class-specific prediction biases. Its compatibility with a wide range of TTA methods adds to its practical value. The experiments provide a comprehensive evaluation of its performance, and the reported results support the effectiveness of this simple approach.

Critical weaknesses

The methodological novelty appears limited. Under a fixed bias-free linear classifier and matched statistics, NEO’s feature centering is equivalent to logit centering. Relative to this operation, LAST’s core transformation mainly adds class-wise scaling by estimated standard deviations.

The authors report extensive experimental gains, but the explanation of why the method works remains insufficient. Propositions 3.1 and 3.2 describe elementary properties of entropy, logit scaling, and standardization. However, the paper does not sufficiently connect these properties to the claim that “Overly attenuated logit scales weaken class discrimination and degrade adaptation reliability” (Lines 244–245), or explain how they account for LAST’s effectiveness in mitigating collapse.

I also used an AI agent to statically inspect the provided code and identified a significant inconsistency in the implementation when batch size=1 . Equation (5) estimates the temporal scale using the deviation of the current logits from the previous EMA mean. In contrast, the code directly sets the batch standard deviation to zero when the batch size is one.

if logits.size(0) == 1:
    batch_mean = logits.squeeze(0).detach()
    batch_std = torch.zeros_like(batch_mean).detach()

The code then applies the EMA update from Equation (4), causing the running standard deviation to decay geometrically, independently of the observed logit fluctuations. This differs substantially from Equation (5) and raises concerns about the reproducibility of the batch-size-one results in Table 3.

Questions for authors
  1. Could the authors provide a controlled comparison of NEO-style centering, mean subtraction only, standard deviation scaling only, and full LAST? It would be helpful to evaluate these variants under matched settings, both as output post-processing and with their outputs used to compute the adaptation loss. These ablations would help identify which components drive the improvements.
  2. As discussed in the weaknesses, were the results in Table 3 obtained using the code provided in the supplementary material, or a different implementation? If a different version was used, could the authors provide the corresponding code and configuration?
Reviewer_AMV2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LAST, which standardizes class-wise logits using exponential moving averages of test steam statistics to improve calibration and mitigate TTA collapse. It supports standalone BP-free use, output calibration for existing methods, and integration into the adaptation loss as LAST$^\dagger$.

Critical strengths
  • The method is simple and readily integrates with existing TTA methods without introducing additional trainable parameters or changing the model architecture.

  • Experiments report accuracy gains across multiple benchmarks and both BP-based and BP-free methods, with substantial improvements from LAST$^\dagger$ in continual adaptation setting.

  • Ablation study is useful because it provides empirical support for the design choices.

Critical weaknesses
  • The causal evidence that logit norm drift is a key driver of collapse is insufficient. LAST changes not only logit norms but also class-wise biases, scales, and gradients, making it difficult to determine which effects primarily account for the performance gains.

  • Intuitively, class-wise standardization may nudge the model toward a more balanced distribution of predicted classes, potentially suppressing useful score differences reflecting the true class prior. However, no evaluation under global class imbalance is reported.

  • Since collapse in TTA can depend on the random seed, evaluation across multiple runs is particularly important. However, per-seed performance and collapse rates are not reported.

  • The runtime and memory overhead of LAST$^\dagger$ is not reported. It is also unclear whether gradients flow through the statistics computation, making the actual update procedure and its cost difficult to access.

Questions for authors
  • Could the authors compare mean-only, std-only, full LAST, and sample-wise norm under the same BP setup?

  • Could the authors evaluate LAST and LAST$^\dagger$ under varying degrees of global class imbalance?

  • What are the mean and standard deviation of accuracy and the collapse rates across seeds for each baseline with and without LAST and LAST$^\dagger$?

  • What runtime and memory overhead does LAST$^\dagger$ incur, and is stop-gradient applied to its statistics?

Reviewer_3nPGOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LAST addresses failure during test-time adaptation by standardizing each class's logit channel using running means and scales. It can calibrate predictions alone or enter an existing adaptation loss. The paper links unstable logit scale to changes in confidence and reports improvements across four benchmarks. Its most striking continual results come from integrating LAST into backpropagation, rather than correcting outputs after the model has collapsed.

Critical strengths
  • The proposed module has useful empirical support. Table 1, reports ImageNet-C accuracy of 55.5 for source inference and 60.7 for standalone LAST; TENT rises from 58.3 to 65.9 with output calibration.

  • The continual experiments distinguish output correction from optimization stabilization. Table 2: TENT averages 29.9 without LAST, 34.2 with output-only LAST, and 66.4 with BP-integrated LAST. DeYO improves from 45.6 to 64.4 with integration.

Critical weaknesses
  • The explanation for why LAST prevents collapse is less convincing than the empirical results. The analysis shows how logit scaling affects confidence, but does not establish how it leads to collapse during adaptation. LAST also changes the relative scores of different classes and, when integrated into backpropagation, the resulting parameter updates. Its improvements therefore do not isolate the role of logit-scale control. The mean-only and standard-deviation-only ablations in Table 5 are useful, but do not separate these effects on collapse during continual adaptation. Comparing scalar scale control, class-wise centering, and full LAST within the same continual adaptation setting would strengthen the proposed explanation.

  • The manuscript does not quantify uncertainty in small accuracy gains. This limits conclusions about sub-point gains. It does not establish that they are nonsignificant or negate the large continual recoveries.

Questions for authors
  • Can an existing-setting control separate scalar norm regulation from class-wise mean correction when measuring collapse?
  • What invariance conditions apply to Algorithm 1 with fixed initialization, EMA history, and epsilon? Please distinguish exact and approximate statements.
  • Are the current-batch statistics detached during the BP-integrated loss? Please state the gradient convention explicitly
  • Please give seed counts or run variability used in table 1
  • Please consider citing and discussing ElAraby, Nashed and Paull, TOOD: Task-Aware Out-of-Distribution Score Calibration for Continual Learners [1], in Section 2. Both methods address drift in output statistics. This is a related-work suggestion, with no rating penalty or required new experiment.

References:

1 - ElAraby, Mostafa, Samer B. Nashed, and Liam Paull. "TOOD: Task-Aware Out-of-Distribution Score Calibration for Continual Learners." arXiv preprint arXiv:2607.29592 (2026).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.