acceptodds
Under review as a conference paper at ICLR 2027

Environment-Conditioned Tail Reweighting for Invariant Learning under Mixed Shifts

Abstract

Out-of-distribution generalization becomes challenging when spurious correlations vary across environments while difficult or underrepresented samples remain insufficiently emphasized within them. Invariant learning and reweighting address complementary aspects of this mixed-shift problem, yet applying reweighting only to prediction leaves the invariance constraint evaluated on a different training risk. We introduce Environment-Conditioned Tail Reweighting (ECTR), which couples cross-environment TV invariance with within-environment adversarial tail weighting through a shared weighted risk. Sample weights are normalized within each environment, and the resulting risk is used consistently for both predictive learning and the total-variation (TV) stationarity penalty, while an environment-wise KL term controls adversarial concentration. Our analysis characterizes the stationarity-dependent weighting signal and its KL-regularized distributional interpretation, alongside conditional-risk and optimization properties under the stated assumptions. Controlled mixed-shift ablations consistently support the shared-risk coupling across three difficulty settings. Across synthetic and real-world benchmarks, ECTR achieves favorable results under both observed- and inferred-environment settings, including improvements over TV-based parent methods and strong performance against additional invariant-learning baselines. Training-corruption experiments further show that ECTR assigns less excess weight to corrupted samples than the tested fixed-tail rules. Together, these results provide a unified theoretical and empirical account of environment-conditioned tail reweighting for invariant learning under mixed shifts.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Invariant Risk Minimization, Total Variation, OOD Generalization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Jwe6OpenReview ↗
Rating
2
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ECTR for settings where spurious correlations change across environments and difficult examples receive too little attention within each environment. Its main idea is to learn sample weights and use the same weighted risk for prediction and the TV stationarity penalty. Weights are normalized within each environment, and a KL penalty limits their concentration. The paper analyzes the resulting weighting signal and evaluates the method through mixed-shift ablations, benchmark comparisons, and studies of corruption and environment information.

Critical strengths
  1. The shared-risk design is well motivated. I find it sensible to evaluate stationarity on the same weighted risk used for prediction. Appendix A.4 makes the effect on the adversary update explicit. The Main-only ablation directly tests this choice. I consider this the paper's strongest result.

  2. The additional controls help explain the results. I appreciate the comparison using identical frozen environment memberships for ECTR and TV-IRM , which helps separate the objective's benefit from environment inference. The corruption study also measures where the sample weight goes. This supports the claim that ECTR places less excess weight on corrupted examples than these fixed-tail rules.

Critical weaknesses
  1. A shared weighted risk need not preserve the invariant prediction target. I understand the motivation for aligning prediction and stationarity. My main concern is that weights depend on the full input, including spurious features, so they can change the conditional label distribution given the invariant features. Preserving the outer environment coefficients does not prevent this change. For a simple squared-loss example, let the input be (C,S)(C,S), with independent, uniform C∈{1,2}C\in\{1,2\} and S∈{−1,+1}S\in\{-1,+1\}. Set Y=C+SY=C+S in one environment and Y=C−SY=C-S in the other. Both environments have E[Y∣C]=C\mathbb E[Y\mid C]=C, and the invariant predictor f(C,S)=Cf(C,S)=C has zero unweighted probe in both. A single input-only scorer can give S=+1S=+1 three times the weight of S=−1S=-1. After normalization, these groups receive mass 0.750.75 and 0.250.25 in each environment. The weighted conditional means become C+0.5C+0.5 and C−0.5C-0.5, and the probes at the same predictor become −1.5-1.5 and +1.5+1.5. The KL cost relative to uniform weighting is finite, about 0.1310.131. All samples initially have squared loss 1, so this violation is introduced by the stationarity-aware reweighting itself. This does not establish that ECTR fails in the reported tasks, but it shows why sharing a risk is not sufficient to justify the new stationarity target. KL controls the amount of reweighting; it does not require preservation of the invariant relationship. I would appreciate conditions or a controlled analysis explaining when this mechanism exposes useful tail variation without penalizing a valid invariant predictor. The current bounds, distributional formulas, and optimization results do not resolve this issue for me.

  2. The benefit of using stationarity to learn the weights is not isolated. Full and Main-only differ in two ways: which risk is tested for stationarity, and whether TV contributes to the adversary's update. Their comparison supports the combined change, but does not distinguish the value of sharing the risk from the value of feeding stationarity back into weight learning. Appendix A.4 shows that the update directions differ; it does not establish that the additional direction favors examples useful for invariant prediction. I would find an aligned loss-only adversary particularly informative: learn weights from Rmain−β KLenvR_{\mathrm{main}}-\beta\,\mathrm{KL}_{\mathrm{env}}, then use those weights in both prediction and TV, without a TV gradient in the weight update. This retains the shared risk and removes only the stationarity feedback to the adversary. The Top-α\alpha and CVaR controls use different selection rules and do not provide this comparison. Without it, I am not yet convinced that stationarity-dependent weight learning is responsible for the gains over simpler aligned reweighting. This distinction matters because that feedback can also alter the invariant target, as illustrated above.

Questions for authors
  1. Under what conditions does the input-dependent reweighting preserve the invariant conditional prediction rule? How would ECTR behave in a controlled example such as W1, where a valid invariant predictor is stationary before reweighting but not afterwards?

  2. Could you compare Full with a loss-only KL adversary whose weights are used in both prediction and TV, with TV excluded only from the adversary update? Using the same architecture and validation budget would help determine the contribution of stationarity-dependent weight learning.

4 comments
AuthorsOpenReview ↗

Q1 W1. Conditional target preservation and finite-support optimization

We thank the reviewer for their careful reading and thoughtful comments. The revision gives an exact criterion for preservation of a shared squared-loss conditional-mean target. For a nonnegative weight r(X)r(X) with finite positive normalization and the stated conditional moments,

=\frac{\mathbb{E}_e[r(X)(Y-\mu(Z))\mid Z]} {\mathbb{E}_e[r(X)\mid Z]}.$$ The denominator is positive, so weighted residual orthogonality is necessary and sufficient. We give sufficient conditions $r(X)=h(Z)$ and $r(X)\perp Y\mid Z,e$, and the equivalent zero-conditional-covariance characterization when the relevant moments exist. The proof is in Appendix C (p. 20), with the statement in Proposition 3 (p. 6). Fix $Z=h(X)$ as the feature variable defining the common target; weights may use the full input. For observed-environment empirical distributions, $r_\theta(X)=\exp s_\theta(X)$ gives ECTR's conditional weights after within-environment normalization. At $f_\mu(X)=\mu(Z)$, with finite reweighted second moments, the squared-loss probe is $$a_e^\mu(r)=-2\mathbb{E}_{e,r}[\mu(Z)\delta_e(Z)],\qquad \delta_e(Z)=\mathbb{E}_{e,r}[Y\mid Z]-\mu(Z).$$ This probe measures a weighted moment of conditional-target displacement. The finite-support study uses $X=(C,S)$, $Z=C$ and $\mu(Z)=C$, and reports $D_\mu$ alongside TV evaluated at each current predictor. Section 4.4 and Appendix C.1 connect these measurements to the preservation criterion. The finite-support model enumerates $C\in\{1,2\}$, $S\in\{-1,+1\}$, $Y=C+eS$ in two environments. Prescribed $3:1$ weights shift the conditional means to $C\pm0.5$ and produce probes $(-1.5,+1.5)$, $\mathrm{TV}=1.5$ and $\mathrm{KL}\approx0.1308$. The squared risk at fixed $f=C$ is one. These analytic quantities are given in Table 7 (p. 20). With a fixed predictor, Full's mean target displacement moves from **0.0380** to **0.8959**, while Aligned ends at **0.0372**. Joint optimization gives mean displacement **0.7004** for Full and **0.0372** for Aligned, with original-measure risks **1.2873** and **1.0039**. The complete TV, KL and uncertainty summaries appear in Table 9 (p. 21) and Table 8 (p. 21). ECTR's shared risk unifies the conditional measure used by the predictor and the stationarity constraint. Preservation of the original conditional prediction target is governed by Proposition 3's residual-orthogonality condition. Full-input weights can violate this condition, and the finite-support study quantifies the resulting target displacement and reference-risk changes. The paper specifies these target-preservation conditions and uses matched ablations and prediction interventions to evaluate the feedback mechanism on the studied tasks.
AuthorsOpenReview ↗

Q2 W2. Contributions of risk alignment and TV feedback

We thank the reviewer for their careful reading and thoughtful comments. Aligned loss-only isolates the feedback contribution. Both Aligned and Full use the conditional weighted risk in prediction and TV. Aligned's weight direction is gθAligned=∇θRmain−β∇θKLenvg_\theta^{\mathrm{Aligned}}=\nabla_\theta R_{\mathrm{main}}-\beta\nabla_\theta\mathrm{KL}_{\mathrm{env}}. At the same parameter point (θ,Φ,Ψ)(\theta,\Phi,\Psi), with fixed environment memberships and training time tt, the selected sample-adversary directions satisfy

=\frac{\lambda_{\mathrm{eff},t}}{E}\sum_e\operatorname{sign}_0(a_e) \sum_{i\in\mathcal I_e}b_i\nabla_\theta\pi_\theta(i\mid e),$$ where $b_i=\left.\nabla_w\ell_i(w)\right|_{w=1}$ and $\operatorname{sign}_0(0)=0$. The predictor uses the same form of shared weighted risk in both methods. The added direction passes stationarity contributions to the sample adversary; Section 3.2 and Appendix A.4 connect this term directly to the mechanism comparison. The matched four-method comparison uses common hyperparameters, backbone, budget, source-validation selection and five seeds. Full has the highest mean overall, worst-environment and worst-group accuracy among the four methods on all three evaluated datasets; Full raises worst-group accuracy from **57.28% to 58.48%**, **36.62% to 38.28%**, and **71.39% to 72.46%**, respectively (Table 2 (p. 7)). Three complementary measurements support this mechanism. The ESS study retains gains at comparable weight concentration (Table 3 (p. 8)); the frozen-predictor study measures lower spurious-feature sensitivity with higher original OOD accuracy (Table 6 (p. 9)); and the label-noise study measures higher noisy accuracy with NMR near one (Table 16 (p. 26), Table 18 (p. 27)). Together they connect the update-level feedback term to concentration-controlled prediction, intervention sensitivity and robustness to corrupted training labels.

Thank you for the additional analyses. My first question was answered: the authors identified when reweighting preserves the prediction target and showed that Full can shift it in the controlled experiment. The Aligned ablation resolved my second concern by isolating TV feedback and showing its benefit on the evaluated tasks. My overall rating remains unchanged while rise the Informativeness.

Thank you for your follow-up and for confirming that our prediction-target analysis answers your first question, and that the Aligned ablation resolves your second concern by isolating TV feedback and demonstrating its benefit on the evaluated tasks.

We have revised the manuscript in response to feedback from all three reviewers. The revision incorporates the prediction-target analysis, matched mechanism ablations, controls at comparable effective sample sizes, and broader evaluation covering independently varied correlation and variance shifts and four DomainBed benchmarks. These additions clarify the theoretical scope of ECTR and provide further evidence for the contribution of shared-risk alignment and stationarity feedback.

We would appreciate your reassessment of the paper as a whole in light of the revised manuscript and the additional evidence. We note that your overall recommendation remains Weak Reject. Could you clarify which substantive concerns remain decisive for this recommendation? We would welcome the opportunity to address those specific points directly during the discussion period.

Reviewer_PKzxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Environment-Conditioned Tail Reweighting (ECTR) to improve OOD generalization under mixed correlation and diversity shifts. ECTR learns adversarial sample weights normalized within each environment, applying the resulting shared weighted risk to both prediction and TV stationarity. An environment-wise KL regularizer controls weight concentration, while auxiliary information enables learning without observed environment labels. Experiments on synthetic and real-world benchmarks evaluate generalization performance, training-corruption robustness, and the contributions of key design components.

Critical strengths
  1. The paper addresses the practically relevant challenge of OOD generalization under mixed distribution shifts. The proposed shared-risk formulation is well motivated and provides a coherent way to integrate environment-conditioned adversarial reweighting with TV stationarity.

  2. The evaluation covers diverse benchmarks under both observed and inferred environment settings, including training corruption. Controlled ablations clarify key design choices, while sensitivity analyses provide insights into training stability.

Critical weaknesses
  1. Limited methodological novelty. Prior work has explored combining sample reweighting with invariant learning. For example, WRI [1] investigates weighted risk invariance, RSW [2] jointly optimizes predictive loss and an IRM gradient penalty using shared weighted risks, and Zhao et al. [3] leverage an IRM-based criterion to guide sample reweighting. The authors should further clarify the key technical distinctions between ECTR and these methods, as well as its specific contributions.

  2. Insufficient theoretical justification for the proposed invariance mechanism. The paper claims that coupling within-environment adversarial reweighting with TV stationarity facilitates invariant feature learning and improves OOD generalization. However, the theoretical analysis primarily characterizes properties of the objective and optimization dynamics, without explaining why this coupling effectively reduces reliance on spurious features. Specifically, TV stationarity only constrains the first-order derivative of the weighted risk with respect to an auxiliary scalar, which does not directly guarantee reduced reliance on spurious correlations or improved OOD generalization.

  3. In Table 2, Full ECTR improves both worst-group accuracy and the ESS ratio compared with Main-only. However, the current experiments do not clearly distinguish whether the performance gains stem from improved invariance learning or reduced sample-weight concentration. ESS-matched ablations would help isolate the contribution of shared-risk coupling.

[1] Weighted Risk Invariance: Domain Generalization under Invariant Feature Shift. TMLR2024

[2] Covariate-Shift Generalization via Random Sample Weighting. AAAI2023

[3] Enhancing Fairness through Reweighting: A Path to Attain the Sufficiency Rule. ECAI2024

Questions for authors
  1. What are the key technical distinctions between ECTR and prior reweighting-based invariant learning methods.

  2. Can the authors provide further theoretical justification for how shared-risk coupling reduces reliance on spurious features and improves OOD generalization?

  3. Do the worst-group accuracy gains of Full ECTR over Main-only in Table 2 persist when controlling for sample-weight concentration through ESS-matched ablations?

3 comments
AuthorsOpenReview ↗

Q1 W1. Relation to WRI, RSW and bilevel reweighting

ECTR learns environment-conditioned adversarial weights and uses the resulting risk for prediction and TV stationarity. Tail reweighting is soft, adaptive emphasis driven jointly by predictive loss and signed stationarity contributions; environment-wise KL controls concentration around the reference distribution.

WRI constructs cross-density-weighted risks and penalizes disagreement of their values while retaining ERM prediction. RSW samples input-dependent weights to create heterogeneous pseudo-environments and trains shared weighted risks. Zhao et al. (2024), Enhancing Fairness through Reweighting: A Path to Attain the Sufficiency Rule (ECAI 2024), optimize an outer sensitive-group IRM criterion through an inner weighted-ERM response. ECTR directly maximizes the current predictor's loss–TV objective within each environment. The manuscript describes these differences in objective, update direction, response path and normalization in Section 2 (pp. 2 and 3) and Appendix D.2 (p. 22).

The matched controls test the predictive contribution of risk alignment and TV feedback within the ECTR construction. Full raises mean worst-group accuracy from 57.28% to 58.48% on Medium, 36.62% to 38.28% on Hard, and 71.39% to 72.46% on CelebA, using the displayed values in Table 2 (p. 7).

Internal risk-alignment and feedback controls — Table 2 (p. 7). Worst-group accuracy (%), mean ±\pm sample SD. Medium and Hard denote LOGIT-Z settings.

DatasetUniform-TVMain-onlyAlignedFull ECTR
Medium56.38±0.8356.38 \pm 0.8356.21±1.9656.21 \pm 1.9657.28±2.5357.28 \pm 2.5358.48±2.4658.48 \pm 2.46
Hard35.22±0.8635.22 \pm 0.8635.04±1.6235.04 \pm 1.6236.62±2.2536.62 \pm 2.2538.28±2.0938.28 \pm 2.09
CelebA69.73±1.4469.73 \pm 1.4470.23±3.8070.23 \pm 3.8071.39±2.9171.39 \pm 2.9172.46±2.9772.46 \pm 2.97
AuthorsOpenReview ↗

Q2 W2. Stationarity feedback, target conditions and prediction sensitivity

At the same parameter point (θ,Φ,Ψ)(\theta,\Phi,\Psi), with fixed environment memberships and training time tt, Full's weight direction contains the selected TV feedback term λeff,t∂θPTV\lambda_{\mathrm{eff},t}\partial_\theta P_{\mathrm{TV}} in addition to prediction and KL, with sign⁡0(0)=0\operatorname{sign}_0(0)=0. The predictor then learns on that same conditional weighted risk. Proposition 3 (p. 6) characterizes conditional-mean preservation by the weighted residual condition Ee[r(X)(Y−μ(Z))∣Z]=0\mathbb{E}_e[r(X)(Y-\mu(Z))\mid Z]=0, under the stated integrability and positive-normalization assumptions.

Frozen-predictor interventions directly evaluate sensitivity to the spurious feature. Relative to Aligned, Full reduces the mean flip rate by 16.14 percentage points on Hard and 7.26 percentage points on Medium, while attaining higher original OOD accuracy. Its probability changes are 43.50 and 32.46, compared with 52.84 and 41.35 for Aligned. Probability Change is 100 times the mean absolute positive-class probability difference. The intervention reverses the spurious latent mean while preserving the original noise realizations and labels, using the same frozen predictor for each pair; Appendix E.5 specifies the input mapping. On the OOD test distribution it restores positive spurious correlation. Original/counterfactual accuracy, probability change and flips are reported together in Table 6 (p. 9), with training-sample results in Table 21 (p. 29).

The CelebA label-noise study supplies a complementary prediction/weighting measurement. Under 10% training-label corruption, Full reaches 69.27±0.66%69.27 \pm 0.66\% worst-group accuracy, compared with 64.81±2.26%64.81 \pm 2.26\% for Aligned and 62.78±0.94%62.78 \pm 0.94\% for Main-only. Its two environment NMR means, 0.992 and 0.995, are close to proportional corrupted-label mass. See Table 16 (p. 26) and Table 18 (p. 27).

Frozen-predictor OOD sensitivity — Table 6 (p. 9). Five-seed mean ±\pm sample SD. Accuracies and flip rates are percentages; probability change is 100100 times the mean absolute positive-class probability difference. CF denotes the counterfactual input.

DataMethodOriginal acc.CF acc.Prob. changeFlip rate
HardUniform-TV44.92±0.8044.92 \pm 0.8075.03±0.8175.03 \pm 0.8159.21±4.8059.21 \pm 4.8050.54±2.1750.54 \pm 2.17
HardMain-only44.86±1.1044.86 \pm 1.1078.56±1.1478.56 \pm 1.1461.69±5.3261.69 \pm 5.3251.91±1.3951.91 \pm 1.39
HardAligned46.41±1.4146.41 \pm 1.4171.29±0.9771.29 \pm 0.9752.84±7.4152.84 \pm 7.4142.62±1.5842.62 \pm 1.58
HardFull48.31±1.3848.31 \pm 1.3867.53±0.6967.53 \pm 0.6943.50±2.3943.50 \pm 2.3926.48±1.4426.48 \pm 1.44
MediumUniform-TV65.24±0.5165.24 \pm 0.5181.37±1.0281.37 \pm 1.0248.93±5.7048.93 \pm 5.7032.09±1.5632.09 \pm 1.56
MediumMain-only64.91±1.1864.91 \pm 1.1883.54±1.5483.54 \pm 1.5450.41±6.1450.41 \pm 6.1435.14±0.9235.14 \pm 0.92
MediumAligned66.50±1.8366.50 \pm 1.8379.25±1.2679.25 \pm 1.2641.35±4.9541.35 \pm 4.9526.53±0.8726.53 \pm 0.87
MediumFull69.94±1.4569.94 \pm 1.4574.31±1.1074.31 \pm 1.1032.46±3.7332.46 \pm 3.7319.27±0.8919.27 \pm 0.89

Noise mass ratio under label corruption — Table 18 (p. 27). Mean ±\pm sample SD. Environments e0/e1 are the high-/lower-alignment training environments; NMR of one is proportional corrupted-label mass.

MethodNMR, e0NMR, e1
Main-only1.0367±0.0311.0367 \pm 0.0311.0380±0.0241.0380 \pm 0.024
Aligned1.0151±0.0291.0151 \pm 0.0291.0160±0.0201.0160 \pm 0.020
Full0.992±0.0170.992 \pm 0.0170.995±0.0220.995 \pm 0.022
AuthorsOpenReview ↗

Q3 W3. Predictive gains at comparable ESS

We thank the reviewer for their careful reading and thoughtful comments. The ESS-controlled study compares methods with similar attained ESS within each of three concentration levels. Actual ESS is reported next to each prediction metric. Full's mean worst-group gains over Main-only are 7.09, 9.40 and 7.19 percentage points across the three levels; the corresponding gains over Aligned are 4.62, 4.80 and 4.43 percentage points. Thus the Full advantage persists across the attained concentration levels. The main comparison is Table 3 (p. 8), and complete prediction metrics appear in Table 11 (p. 23).

Prediction at comparable ESS — Table 3 (p. 8). Mean ±\pm sample SD. Target and actual ESS are ratios.

Target ESS ratioMethodWorst-group acc. (%)Actual ESS ratio
0.3Main-only50.624±2.37650.624 \pm 2.3760.311±0.0180.311 \pm 0.018
0.3Aligned53.094±3.29053.094 \pm 3.2900.304±0.0200.304 \pm 0.020
0.3Full57.714±1.48957.714 \pm 1.4890.298±0.0280.298 \pm 0.028
0.5Main-only49.220±2.65349.220 \pm 2.6530.502±0.0090.502 \pm 0.009
0.5Aligned53.818±2.15253.818 \pm 2.1520.492±0.0150.492 \pm 0.015
0.5Full58.616±2.25758.616 \pm 2.2570.493±0.0350.493 \pm 0.035
0.7Main-only51.734±1.55651.734 \pm 1.5560.716±0.0050.716 \pm 0.005
0.7Aligned54.495±2.76654.495 \pm 2.7660.708±0.0100.708 \pm 0.010
0.7Full58.928±1.97458.928 \pm 1.9740.701±0.0320.701 \pm 0.032
Reviewer_ydgmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses mixed shifts in OOD generalization, involving changes in spurious correlations across environments and underrepresented variations within environments. The authors propose Environment-Conditioned Tail Reweighting (ECTR), which combines adversarial sample reweighting with TV-based invariant learning. ECTR uses the same environment-conditioned weighted risk for prediction and stationarity, with KL regularization to control weight concentration. Experiments on eight datasets, including synthetic and real-world datasets, demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The paper addresses correlation and diversity shifts, which are important challenges in OOD generalization. Combining environment-conditioned reweighting with invariant learning through a shared weighted risk is a reasonable approach to handling these shifts.

  2. The authors provide comparisons with various invariant learning and robust learning baselines under both observed- and inferred-environment settings. The ablation studies also demonstrate the empirical benefits of shared-risk coupling and KL regularization.

Critical weaknesses
  1. Limited evaluation on standard OOD generalization benchmarks. Although the paper evaluates ECTR on multiple datasets, the experiments largely rely on dataset-specific evaluation protocols. The lack of comprehensive comparisons under standardized benchmarks, such as DomainBed [1] or WILDS [2], makes it difficult to assess the generalizability of the reported improvements across established OOD evaluation settings.

  2. Limited analysis of individual shifts within mixed-shift settings. In the controlled mixed-shift experiments (LOGIT-Z) , correlation and diversity shifts are varied simultaneously, making it difficult to distinguish their individual effects. As it is controlled experiments, independently varies each shift under matched conditions would help clarify how ECTR addresses the two challenges.

  3. Sensitivity to auxiliary environment information. The inferred-environment setting relies on auxiliary variables that may be strongly associated with environment membership (e.g., construction year for House Price). As shown in Table 15, ECTR's performance on House Price degrades substantially when auxiliary information is shuffled or removed. This raises concerns about its dependence on informative auxiliary variables and its generalizability when such information is unavailable or weakly correlated with the underlying environments.

[1] Gulrajani, Ishaan, and David Lopez-Paz. "In search of lost domain generalization." arXiv preprint arXiv:2007.01434 (2020).

[2] Koh, Pang Wei, et al. "Wilds: A benchmark of in-the-wild distribution shifts." International conference on machine learning. PMLR, 2021.

Questions for authors

Please address the weaknesses.

6 comments
AuthorsOpenReview ↗

W1. Standardized domain-generalization evaluation

We thank the reviewer for their careful reading and thoughtful comments. We extended ECTR to PACS, OfficeHome, VLCS and TerraIncognita under the standard DomainBed evaluation protocol.

ECTR achieves average accuracies of 86.0%, 67.0%, 79.0% and 48.4%, respectively, using the one-decimal display in the revision. These means exceed the reported ERM, GroupDRO, IRM and V-REx baselines on all four datasets. The four baselines cover empirical risk minimization, invariant learning, risk-variance regularization, and environment-level robust optimization. The baseline rows reproduce published DomainBed results, and ECTR uses the corresponding evaluation and model-selection protocol. The manuscript gives the average comparison in Table 1 (p. 7) and the complete domain-level results in Table 10 (p. 23); protocol details appear in Section 5.1 (p. 7) and Appendix E.1 (p. 22).

Domain-generalization accuracy (%) — Table 1 (p. 7). Reported arithmetic mean over four target domains.

MethodPACSOfficeHomeVLCSTerra
ERM85.5±0.285.5 \pm 0.266.5±0.366.5 \pm 0.377.5±0.477.5 \pm 0.446.1±1.846.1 \pm 1.8
GroupDRO84.4±0.884.4 \pm 0.866.0±0.766.0 \pm 0.776.7±0.676.7 \pm 0.643.2±1.143.2 \pm 1.1
IRM83.5±0.883.5 \pm 0.864.3±2.264.3 \pm 2.278.5±0.578.5 \pm 0.547.6±0.847.6 \pm 0.8
V-REx84.9±0.684.9 \pm 0.666.4±0.666.4 \pm 0.678.3±0.278.3 \pm 0.246.4±0.646.4 \pm 0.6
TV-IRM83.8±0.583.8 \pm 0.566.8±0.466.8 \pm 0.478.5±0.378.5 \pm 0.345.7±1.045.7 \pm 1.0
ECTR86.0±0.686.0 \pm 0.667.0±0.367.0 \pm 0.379.0±0.479.0 \pm 0.448.4±0.948.4 \pm 0.9
AuthorsOpenReview ↗

W2. Independent correlation and diversity effects

We thank the reviewer for their careful reading and thoughtful comments. We independently trained Uniform-TV, Main-only, Aligned loss-only and Full ECTR on newly generated LOGIT-Z data. Each frozen predictor is evaluated under four paired conditions: source (C0D0C_0D_0), correlation reversal (C1D0C_1D_0), increased feature-noise variance (C0D1C_0D_1), and their combination (C1D1C_1D_1). The paired conditions share sample IDs, labels, invariant latent variables and base random draws. Appendix E.2 (pp. 23 and 24) provides the complete training and evaluation protocol.

Four-condition accuracies excerpted from Table 4 (p. 8).

MethodC0D0C_0D_0 (%)C1D0C_1D_0 (%)C0D1C_0D_1 (%)C1D1C_1D_1 (%)
Uniform-TV82.37±0.0382.37 \pm 0.0370.58±0.8070.58 \pm 0.8080.04±0.4780.04 \pm 0.4761.36±1.2361.36 \pm 1.23
Main-only82.05±0.0482.05 \pm 0.0471.64±0.0671.64 \pm 0.0680.06±0.2280.06 \pm 0.2261.42±0.8161.42 \pm 0.81
Aligned loss-only81.46±0.0381.46 \pm 0.0373.61±0.0773.61 \pm 0.0779.57±0.2479.57 \pm 0.2462.71±0.9462.71 \pm 0.94
Full ECTR80.84±0.0280.84 \pm 0.0275.14±0.0675.14 \pm 0.0678.90±0.2178.90 \pm 0.2164.43±1.0964.43 \pm 1.09

Accuracy is mean ±\pm sample SD over five training seeds (ddof=1\mathrm{ddof}=1).

Full achieves the highest mean accuracy under correlation reversal and mixed shift, exceeding Aligned by 1.53 and 1.72 percentage points. The correlation-only loss decreases from 11.79 percentage points for Uniform-TV to 10.41, 7.85 and 5.70 percentage points for Main-only, Aligned and Full. Aligned and Full have similar variance-only losses (1.89 and 1.94 percentage points), locating the feedback benefit in robustness to correlation reversal. Their source accuracies, 81.46% and 80.84%, characterize the tradeoff between fitting source associations and maintaining accuracy after reversal. Full's source-to-mixed loss is 16.41 percentage points, compared with 18.75 percentage points for Aligned. See Table 4 (p. 8) and Appendix E.2 (pp. 23 and 24).

AuthorsOpenReview ↗

W3. Auxiliary information and fixed memberships

We thank the reviewer for their careful reading and thoughtful comments. The original House temporal-regression study shows that construction year provides useful environment information for this task: setting the auxiliary input to zero increases ECTR's worst-environment MSE from 0.5404 to 0.8471 (Table 14 (p. 25)). The added classification studies examine information degradation across different tasks and compare learning objectives under identical fixed memberships.

We evaluate Full and matched Uniform-TV on Adult-E4 and ten-class, three-color CMNIST with five seeds for every method, dataset and condition. Appendix B.2 defines their tasks, input features, environments and evaluation groups. The six settings are full auxiliary information, zero auxiliary information, permutation of 50% of sample rows, and fixed oracle, uniform or soft-0.5 memberships. Fixed-membership comparisons provide the same frozen qq to both methods.

Full has higher mean worst-group or worst-environment accuracy in all 12 dataset–condition combinations. With oracle memberships, Adult-E4 gives 81.857±0.190%81.857 \pm 0.190\% for Full and 80.349±0.247%80.349 \pm 0.247\% for Uniform-TV; CMNIST gives 95.768±0.141%95.768 \pm 0.141\% and 94.873±0.303%94.873 \pm 0.303\%. The matched-qq comparison directly evaluates the objective under identical environment information. Zero auxiliary information and uniform memberships yield the same effective conditional weighting in these settings. The main comparison is Table 5 (p. 9); complete metrics appear in Table 12 (p. 24) and Table 13 (p. 25).

Auxiliary-information and fixed-membership comparison — Table 5 (p. 9). Accuracy (%), mean ±\pm sample SD. F: Full ECTR; U: Uniform-TV. Adult-E4 reports worst-group accuracy and CMNIST reports worst-environment accuracy.

ConditionAdult-E4 FAdult-E4 UCMNIST FCMNIST U
Full auxiliary81.762±0.07681.762 \pm 0.07680.797±0.34680.797 \pm 0.34695.595±0.20295.595 \pm 0.20294.825±0.16694.825 \pm 0.166
Partial 50%80.945±0.15880.945 \pm 0.15880.238±0.31980.238 \pm 0.31994.748±0.13994.748 \pm 0.13994.309±0.10194.309 \pm 0.101
Zero auxiliary80.477±0.20080.477 \pm 0.20080.110±0.31180.110 \pm 0.31193.893±0.11493.893 \pm 0.11493.055±0.13393.055 \pm 0.133
Fixed oracle81.857±0.19081.857 \pm 0.19080.349±0.24780.349 \pm 0.24795.768±0.14195.768 \pm 0.14194.873±0.30394.873 \pm 0.303
Fixed soft 0.580.825±0.11080.825 \pm 0.11079.601±0.14779.601 \pm 0.14794.035±0.11594.035 \pm 0.11593.924±0.20893.924 \pm 0.208
Fixed uniform80.477±0.20080.477 \pm 0.20080.110±0.31180.110 \pm 0.31193.893±0.11493.893 \pm 0.11493.055±0.13393.055 \pm 0.133

Comments on the paper

2 comments
AuthorsOpenReview ↗

Title: Follow-up on Our Rebuttal

Dear Reviewer and Area Chair,

Thank you for your thoughtful review. In our rebuttal, we have carefully addressed your comments and concerns, providing additional experimental evidence and theoretical clarifications regarding the shared-risk coupling mechanism.

As the author–reviewer discussion period approaches its November 18 deadline, we would appreciate your feedback on whether our responses have adequately addressed your main concerns or if any questions remain. We would be happy to provide further clarification during the remaining discussion period.

Thank you again for your time and consideration.

Best regards,

The Authors

We thank the reviewers for their detailed feedback. Across our responses to all three reviewers and the follow-up discussion with ydgm, we substantially strengthened the manuscript in four areas: mechanism attribution, standardized evaluation, theoretical clarification, and applicability conditions.

1. A clearer contribution supported by targeted mechanism controls. We clarified ECTR’s relationship to WRI, RSW, and bilevel reweighting, specifying the differences in objectives, normalization, and feedback paths. We introduced the requested Aligned loss-only control to separate shared-risk alignment from TV feedback into weight learning. Full ECTR improves mean worst-group accuracy over Aligned on LOGIT-Z Medium, Hard, and CelebA, and the advantage persists at comparable effective sample sizes. Complementary prediction interventions, label-noise experiments, and independently controlled correlation/variance shifts connect the update mechanism to predictive behavior under the tested shifts.

2. Expanded standardized evaluation and a direct TV-parent comparison. We added DomainBed experiments on PACS, OfficeHome, VLCS, and TerraIncognita. ECTR achieves the highest mean accuracy among the compared methods on all four datasets. In response to ydgm’s follow-up, we additionally reported TV-IRM under the same DomainBed protocol, together with standard errors for the trial-level performance differences. The mean improvement over TV-IRM is approximately 1.4 percentage points across the four datasets, with gains of 2.2 points on PACS and 2.7 on TerraIncognita. These results complement the controlled ablations with direct evidence of the extension’s practical value.

3. An explicit theoretical characterization of target preservation. The revision adds an exact residual-orthogonality criterion for preserving a shared squared-loss conditional-mean target, with a proof and a connection to the scalar stationarity probe. We also analyzed the reviewer’s finite-support construction and evaluated both fixed-predictor and joint optimization. This establishes when target preservation holds and quantifies how input-dependent weighting can change the target, giving the theoretical claims a precise scope.

4. Direct evidence on auxiliary-information dependence. We added information-degradation and identical-membership controls on Adult-E4 and CMNIST, followed by the requested auxiliary-free ERM comparison on House. The House results show lower mean errors with full or partially masked auxiliary information, while ERM performs better when auxiliary information is completely removed. We explicitly identify this limitation of the inferred-environment extension. Fixed-membership comparisons separately support the contribution of ECTR’s learning objective under identical environment information. REBUTTAL_ALIGNED

The revised manuscript integrates the new theoretical analysis and experimental evidence with corresponding sections, tables, and appendices. Together with the follow-up results, these additions provide a clearer account of what ECTR contributes, which components produce the observed gains, and under what information conditions those gains apply.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.