acceptodds
Under review as a conference paper at ICLR 2027

AVERT: W1-Anchored Verification for Reliable Test-Time Adaptation in Regression

Abstract

Test-time adaptation (TTA) for regression lacks the entropy-based confidence signal that classification TTA relies on, so naive adaptation can shift predictions in arbitrary directions. Analyzing four recent regression and time-series TTA methods, we identify four failure modes—negative transfer on low-shift data, unreliable low-order moment matching, over-adaptation, and label-shift confusion, where a near-perfect source model is corrected toward the source prediction distribution, moving predictions away from the target truth (on Concrete, SSA inflates MAE by ). We propose AVERT, a safe TTA framework built on three principles. (1) W1-anchoring: because target labels are standardized with source statistics, the 1-Wasserstein distance from a corrected prediction to the source label distribution is a label-free verification signal for regression TTA (correlation with true MAE change), with a finite-sample concentration analysis. (2) Verification-gated adaptation: a non-parametric quantile-transport bridge (QTB) combined with a gated test-time self-supervised representation adaptation (TT-SRA), both verified and gated by the W1 anchor. (3) No-harm safeguards: a safe-return gate detects label-shift confusion and refuses adaptation. Across 15 regression benchmarks and three backbones, AVERT significantly outperforms Original () and all six adaptation baselines (up to ), achieves the lowest average MAE () with the best average rank on two of three backbones, is the only method that never catastrophically degrades on any of the 45 dataset–backbone combinations, and transfers to a fully external dataset with zero retuning.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
test-time adaptation, regression, label-free verification, Wasserstein distance, quantile transport

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_NbV8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses test-time adaptation for regression. It argues that existing methods fail in four ways, including "label-shift confusion," where aligning target predictions to the source prediction distribution moves them away from the truth. It then proposes AVERT, which uses the 1-Wasserstein distance between predictions and the source label distribution as a label-free verification signal (W1-anchoring). Adaptation is done by quantile transport of predictions or features (QTB) and a masked-reconstruction fine-tuning step (TT-SRA), followed by a safe-return gate, a W1 validation clamp, per-sample blending and an anomaly clamp. Theory consists of a triangle-inequality proxy bound, a DKW-based concentration result and a no-harm statement. Experiments cover 15 tabular and time-series regression datasets, three backbones and three seeds against six baselines, plus one external dataset.

Critical strengths

The safety question is the right one for regression TTA. The tables make clear that unconditional alignment can be far worse than no adaptation on tabular regression, and the paper reports where its own method is conservative or loses. Asking when adaptation should be refused is a useful framing.

Low cost. The method is closed-form for the most part and needs no iterative optimization in its main path, which would matter in deployment if the decision rules were shown to generalize.

Critical weaknesses
  • The verification signal is not supported by the theory or by the evidence, and the method depends on many constants that appear fitted to the evaluation data.

    • Theory: The bound MAE ≤ v + δ + κ defines κ as MAE − W1(ŷ, y^t), so it restates the definition. W1 compares marginal distributions only; a model with imperfect ranking can have small W1 and large MAE, and κ is then the dominant term. κ is never measured. The "bounded gain" guarantee inherits the same term and can be vacuous. The "no-harm" property holds only when the gate returns the original prediction.

    • Assumption: Everything rests on W1(y^t, y^s) being small. The paper calls this "directly checkable per deployment", but checking it requires target labels. In the central example Concrete, the target label mean is about −1.8 source standard deviations, so the assumption fails. What the paper names "label-shift confusion" is ordinary label shift, the case where the anchor is invalid.

    • Circularity: QTB maps target predictions onto the source prediction quantiles (Eq. 2), which for a competent source model are close to the source labels. It therefore lowers v by construction, so v cannot independently verify it. The introduction also states that aligning to source predictions is the root cause of failure, yet this is what Eq. 2 does.

    • Evidence: The r = +0.964 in Figure 1 comes from 15 points, one seed and one backbone. Judging from the plot, most points sit close to the origin and the correlation seems to rest largely on a single distant point. For SSA, PETSA and TAFAS the same correlation is about −0.6: the signal fails on the harmful corrections a verifier most needs to catch. The "natural label shift" study has label drift of only 0.05 to 0.18 σ and the gate never fires, so it does not test the failure regime.

    • Constants: The pipeline has more than twenty hand-set thresholds and blending weights. No selection procedure or validation split is described, and the statement that Airfoil "was not used for hyperparameter selection" implies the 15 benchmarks were. The W1 gate threshold of 2.0 lies between the anchor distances of Concrete and Energy Efficiency: moving it to 1.4 raises Concrete MAE from 0.390 to 0.623, and moving it to 2.6 raises Energy Efficiency from 0.711 to 1.201. On Airfoil the reported range starts at exactly 2.0. This is not threshold insensitivity. One external dataset is not enough to show the configuration transfers.

  • The empirical gains are narrow, the comparison is not convincing, and closely related work is missing.

    • Where the gain comes from: By my rough calculation from the three per-backbone tables, most of the pooled MAE improvement over Original, on the order of three quarters, comes from Energy Efficiency alone. On the MLP backbone AVERT is identical to Original on 5 of 15 datasets and slightly worse on three. Where baselines help most (NWP-s34: about 0.51 vs 1.247 for AVERT; Gas Turbine), AVERT keeps little of the benefit. The result is closer to "rarely adapt" than to reliable adaptation.

    • Baselines: TAFAS and COSA were redesigned by the authors for a setting they were not built for. TENT is run "by entropy minimization" although the paper states entropy is undefined for regression. SSA was proposed for image regression under covariate shift. SSA, PETSA, TAFAS and CORAL produce nearly identical numbers on several datasets (e.g., PGCB 1.93 to 1.95), which suggests they reduce to the same operation here. Catastrophic baseline results on datasets with strong label shift say little about those methods.

    • Setting: The method processes the target set offline and stores the full source label vector and source feature matrix. This is closer to transductive domain adaptation with source statistics than to TTA, yet no source-available regression UDA baseline is included. Target batch size is not stated.

    • Ablation and statistics: Only two ablations are given. TT-SRA, feature-level QTB, the clamps and per-sample blending are never isolated, so the source of the Energy Efficiency gain is unknown. The pooled Wilcoxon test treats 135 dependent observations (seeds and backbones of the same dataset) as paired samples.

    • Related work: W1-anchoring is presented as novel, but COT (Lu et al., NeurIPS 2023) already estimates error without labels from the Wasserstein distance between target predictions and the source label distribution, i.e., with the source label distribution as the reference, as here. Quantile mapping is a long-standing bias-correction technique. Adaptation under target (label) shift (e.g., Zhang et al., ICML 2013), regression UDA (RSD, Chen et al., ICML 2021; DARE-GRAM, Nejjar et al., CVPR 2023) and tabular TTA that handles label shift (AdapTable, Kim et al., 2024; TabLog, ICML 2024; PFT3A, ICLR 2026) are not discussed.

Questions for authors

How were the thresholds and blending constants chosen, and on what data? Was any target-labeled data from the 15 benchmarks used?

How are source and target defined for datasets without a temporal axis (Concrete, Energy Efficiency, California Housing, SGEMM)?

Please report W1(y^t, y^s) for every dataset.

Please report κ for Original and AVERT on each dataset, so the bounds can be judged numerically.

What is the correlation in Figure 1 without Energy Efficiency, as a rank correlation, and over all seeds and backbones?

What are the pooled results without Energy Efficiency? How often does each of the safe-return gate, the W1 clamp and the anomaly clamp trigger per dataset?

Please ablate TT-SRA, feature-level QTB, per-sample blending and the two clamps separately.

What objective does TENT optimize here? How exactly were TAFAS and COSA modified, and do the unmodified methods reproduce their published results in their original settings?

Is the whole target set processed at once? How does AVERT behave with small online batches?

Reviewer_W6cmOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies test-time adaptation (TTA) for regression under distribution shift. The authors identify four failure modes in existing regression and time-series TTA methods: negative transfer from unconditional adaptation, unreliable correction directions from low-order moment matching, over-adaptation and catastrophic degradation, and label-shift confusion. To address these issues, they propose AVERT, which adapts only when the measured shift is sufficient, verifies candidate corrections against a label-free anchor, and limits the correction magnitude. Specifically, AVERT uses the 1-Wasserstein distance between predictions and the source label distribution as a verification signal. It combines a non-parametric quantile transport bridge with gated self-supervised representation adaptation, while using safe-return gating, conservative blending, and correction clamps to reduce harmful updates. Across 15 regression benchmarks and three backbones, AVERT achieves the lowest overall average MAE and the best average rank on each backbone compared with the unadapted model and six adaptation baselines, while reducing the risk of large performance degradations.

Critical strengths
  1. The paper identifies meaningful failure modes in regression TTA that could help future research.
  2. The method includes simple and potentially useful components that could be used into other adaptation methods, even though the components performance need more evidences to support the state-of-the-art claim.
Critical weaknesses
  1. Insufficient failure-mode validation. The experiments demonstrate AVERT’s overall effectiveness but do not establish that its components specifically resolve all four claimed failure modes. This leaves a gap between the paper’s motivation and its methodological contributions. Besides, the statement of all four failure modes should be more precise and need enough evidence to prove they exist and matter.
  2. Limited justification of key design choices. The choice of W₁, the shift-dependent routing thresholds DD, and the maximum-based shift score lack sufficient justification or controlled comparisons. The relationship to prior optimal-transport-based error characterization also needs clarification.
  3. Incomplete experimental evidence. Missing variability estimates, potentially outlier-sensitive correlation, and insufficient evidence of source-model adequacy limit confidence in the reported results and their interpretation.
  4. Poor presentation of important methodological sections. Key quantities are used before being clearly defined, and the roles of different thresholds are not clearly explained, making the method difficult to follow and assess.
  5. Limited novelty. AVERT largely combines existing adaptation mechanisms with heuristic safeguards, while the novelty of W₁-anchoring relative to prior optimal-transport-based approaches is insufficiently established. Also the necessity of W1 is also insufficiently established.
Questions for authors

W₁-Anchoring

  1. In Section 3.2, W₁-anchoring uses the triangle inequality to relate an unavailable target-distribution distance to a computable source-distribution distance. The choice of W₁ appears motivated by MAE as the evaluation metric. Are there other reasons for choosing W₁?
  2. Please clarify and distinguish this idea from Characterizing Out-of-Distribution Error via Optimal Transport.

Results

  1. In Figure 1, please explain the outlier in the lower-left corner and its effect on Pearson’s rr. Please also provide Pearson’s rr and the plot without the outlier.
  2. Please provide standard deviations for all results.

Method Clarifications

  1. In Section 3.3, “Feature-level QTB (high shift),” please define DD before using it.
  2. Why is the threshold 0.250.25? Please clarify the statement “When D>0.25D>0.25, prediction-level transport may be too coarse because the regression head collapses information,” and provide evidence or derivation supporting this threshold.
  3. There are several other thresholds for DD in the paper, including 0.150.15 and 0.30.3. Please explain how these thresholds were defined or selected.
  4. Why is the per-dimension shift score defined as the maximum of two discrepancy measures? Please provide evidence that the maximum is necessary.
  5. What is src_mean?

Source-Model Adequacy

  1. Could the authors provide evidence that the source models are sufficiently trained and achieve meaningful predictive performance before adaptation? In particular, please report held-out source-domain performance against a constant-prediction baseline and a competitive supervised baseline, clarify whether src_mae is measured on training or validation data, and provide convergence diagnostics.

Failure-Mode Validation

  1. Does AVERT specifically address the four claimed failure modes? Overall improvements do not establish that each component addresses its associated failure mode. The label-shift-confusion case and safe-return ablation provide relatively direct evidence; comparable evidence is missing for low-shift negative transfer, unreliable correction from low-order moment matching, and over-adaptation. Please evaluate each corresponding component in settings exhibiting its associated failure mode while keeping the remaining components fixed.
Reviewer_sMPjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AVERT, a test-time adaptation (TTA) framework for regression tasks that aims to reduce performance degradation under distribution shifts. The method uses the 1-Wasserstein distance (W1) between target predictions and the source label distribution as a label-free verification signal, and combines Quantile Transport Bridge (QTB) and Test-Time Self-Supervised Representation Adaptation (TT-SRA) to correct predictions. In addition, the method introduces a Safe-Return Gate, W1 Validation Clamp, and Anomaly Clamp to control adaptation risks. The authors evaluate AVERT on 15 regression benchmarks and three model architectures, with additional ablation studies, robustness analyses, and evaluation on an external dataset.

Critical strengths
  1. This paper addresses the lack of reliable label-free verification signals and the risk of negative transfer in regression TTA, which is a meaningful research problem with practical relevance.

  2. The experiments cover 15 regression benchmarks and three model architectures, including comparisons with multiple baselines, ablation studies, robustness analyses, and evaluation on an external dataset.

Critical weaknesses
  1. The paper does not clearly define the different adaptation scenarios. The similarity between source and target label distributions, the agreement between W1 and actual MAE, and source-model accuracy may lead to different adaptation needs and risks. Although the authors introduce the W1 assumption and Safe-Return Gate, it remains unclear which scenarios are addressed, how they can be identified without target labels, and which strategies should be used. I suggest providing intuitive illustrations of these scenarios and systematically evaluating the method under different conditions.

  2. Section 3.2 establishes the W1 verification signal under certain assumptions and discusses its limitations. However, whether these assumptions hold cannot be directly determined without target labels. Although AVERT introduces the Safe-Return Gate and W1 Validation Clamp, it remains unclear how these mechanisms handle cases where W1 becomes unreliable, or why W1-based decisions can distinguish beneficial from harmful adaptations. I suggest clarifying the connection between the theoretical assumptions and practical verification mechanisms, and evaluating their reliability when these assumptions are violated.

  3. AVERT combines multiple adaptation branches, gates, and blending operations governed by numerous predefined thresholds. For example, the shift score (D) controls adaptation and the choice of QTB and TT-SRA, while Safe-Return combines source prediction error, prediction mean shift, and W1 distance. The paper provides limited justification for these thresholds, the selected conditions, or how they are combined. Existing sensitivity analyses and ablations do not sufficiently establish their robustness and necessity across different distributions, or whether simpler strategies could achieve comparable performance.

Questions for authors

Section 3.1 requires retaining the source feature matrix (Z_s). How much source-domain information needs to be stored, and what is the associated storage overhead?

Reviewer_14dXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes an algorithm for Test time adaptation for regression problems. The main insight is that in regression problems it is possible to verify before correcting the distribution shift.

Critical strengths
  • TTA on regression problems is important and interesting.
Critical weaknesses
  • I am unclear on why standardization of target yty^t using source mean and variance would produce a verification signal. How can one find W1(y^,ys)W1(\hat{y}, y^s) is also unclear to me. Here a target sample does not have a matched source sample so I'm not sure how corresponding outputs can be compared.
  • In W1 distance the labels are matched ignoring the input values which feels odd -- two datasets can have zero W1 in terms of their labels even with drastically different intput -> output mapping.
  • I didn't understand the material in 3.3 starting from eq 2. Can you elaborte this more?
  • The input to the Algorithm 1 looks very different from typical test time adaptation settings: generally the source features, source labels, and domain shifts are not given as inputs. I am also confused about hard coded thresholds in all over the algorithm (e.g., line 6, 16, 18).
  • Results need to include confidence intervals as Table 1 MAE values are closeby. Also MAE needs to be put in the perspective of the true output values, which is unclear in this table.
  • To explain the domain shift if the authors describe the KS test and the results that they used to measure domain shift. If the authors describe the domain shift examples early in the paper (intro) it'll improve the readability of the paper.

Overall I think the paper contains interesting technical concepts, but the readability is low.

Questions for authors

As above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.