acceptodds
Under review as a conference paper at ICLR 2027

Minimal Nuisance, Maximal Fidelity: Latent-State-Conditioned Invariance in Domain Generalisation

Abstract

Generalisation across physical systems remains difficult because domains differ in equipment, operating regimes, and system dynamics. As a result, models often rely on system-specific shortcuts rather than learning transferable process behaviour. Existing invariant representation learning methods typically enforce global domain invariance, or condition only on coarse task labels, potentially erasing useful within-label behavioural structure. We argue that the appropriate target is *latent-state-conditioned nuisance invariance* (LaSNI), where domain-specific variation should be removed only after conditioning on the underlying behavioural state. We formalise this target with , where is the learned representation, the domain, and an unobserved latent behavioural state. Since is unobserved, we propose *Local Behavioural Nuisance Suppression* (LBNS), a source-only surrogate that constructs dynamic same-label neighbourhoods, estimates a local behavioural reference, and suppresses cross-domain variation in the residual rather than the full representation. Across six real-world tasks and 30 leave-domain-out folds, LBNS achieves the strongest aggregate transfer and lowest average domain recoverability across all five objectives. An independent descriptor diagnostic shows that its latent neighbours are more behaviourally coherent than random same-label samples in every fold and dataset, including under cross-domain restriction. These results support conditioning nuisance suppression more finely than the observed class when labels hide heterogeneous behaviour.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Domain Generalisation, Representation Learning, Time Series

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DGdqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper attemps to propose a new concept, latent behavioural state, for domain generalization. ss represents the real variant element under domain shift, rather than the shift of relationship between label yy and domain dd. To align representations conditioned on the same ss, the authors refer to a surrogate that could represent the distribution in latent behavioural state. The optimization function thus focuses on tightening the surrogate cluster range and on tightening the representations within each cluster.

Critical strengths
  1. The concept of within-class variation and domain-specific nuisance provides a meaningful ingisht for domain generalization.
  2. The experiment protocol follows the domainbed guidelines to isolate the influence of hyperparameter selections.
Critical weaknesses

The key weakness lies on unclear definition and time-series data bias. please see my questions for details.

Questions for authors
  1. Unclear definition The behavioural state ss is insufficiently specified: what constitutes a state, what information should be preserved, and how it differs from other invariant concepts (e.g. the "invariant information" concept in "Bayesian Domain Invariant Learning via Posterior Generalization of Parameter Distributions"). The algorithm constructs same-label local references without establishing their relationship to ss. The conceptual distinction from MatchDG’s object-conditioned invariance and approximate matching also needs clarification.

  2. Weak connection to time-series structure LBNS uses generic latent-space neighbourhoods, averaging, and L2L_2 penalties while the only temporal modelling comes from the encoder. In other words, the algorithm works on all kinds of data and currently there is no specific adaptions to time series data. This raises my concern that current experiments are manually selected and LBNS fails on other common datasets. The authors should explain and demonstrate what makes the contribution specific to time series, or substantiate its broader applicability through DomainBed experiments on PACS, VLCS, Office-Home, Terra Incognita, and DomainNet.

  3. Insufficient baseline DIVERSIFY is the only explicitly time-series-specific DG comparison. The remaining comparisons are out-of-date, general objectives, which are orinigally designed for image data. Time-series DG methods and up-to-date approaches are required. If possible, also include well-established baselines such as SWAD MIRO.

  4. Unestablished theoretical premise Proposition 2 assumes that the local reference approximates the behavioural component with small error. However, Neighbourhood tightness does not mean zero-nuisance, leaving a substantial gap between the theoretical assumptions and the implemented estimator.

  5. Inconclusive ablations Table 7 does not clearly isolate locality from RBF weighting, and its paired intervals include zero. Table 8 uses identical coefficients for direct and residual alignment, rather than select hyperparameters independently. Matched controls and fully hyperparameter-selections are required to show the attribute of conditioning granularity and residualisation.

  6. Minor writing issues The abstract and introduction use notations to simplify expresions, but the intended concepts are not clearly explained before. Figure 1 contains text that is extreme small, and the connection between the independence target (top) and algorithmic procedure (bottom) is unclear. The explanation of how tightness loss prevents collapse also needs clarification.

1 comment

We thank the reviewer for the helpful comments.

Unclear definition [...]. Behavioural state ss denotes the underlying process condition, such as operating regime, load or fault severity. We aim to preserve meaningful within-class differences even when they can reveal domain identity (lines 053–061).

We now cite Shen et al. (2023), arXiv:2310.16277, and clarify the distinction (lines 068–071): its domain-invariant information DcD_c has the same distribution across domains, whereas ss need not.

The algorithm constructs [...]. Same-label membership ensures task compatibility, not state identity. The relationship to ss rests on the local-coherence assumption and bounded reference-error model in Proposition 2; the independent descriptor diagnostic supports behavioural comparability.

MatchDG directly regularises representations across inferred object matches. LBNS instead aligns residuals around sample-specific local references, allowing locally explained differences mi−mjm_i-m_j between comparable but non-identical states to remain (lines 127–136).

Weak connection to [...] We clarified the motivation in the introduction (lines 053–061) and Related Work (lines 133–136), and the dataset-selection rationale at lines 315–319.

Cyber-physical systems generate sensor time series in which meaningful differences in load, control or regime may identify the installation. Direct alignment can suppress these differences alongside nuisance; LBNS aligns residuals around sample-specific local references, allowing locally explained differences between comparable states to remain.

Our empirical claims concern cross-system time series generalisation, where substantial OOD challenges remain (Wu et al., 2026). Encoder agnosticism does not imply selective reporting or failure on untested modalities. Image-domain experiments would assess broader applicability beyond our current scope, a boundary now acknowledged in the limitations.

Insufficient baseline [...]. We added SWAD, CCIL (sensor time series activity recognition), and FIXED (a general DG method evaluated on time series). ERM, DANN, CCA and CDANN isolate no suppression, global suppression and class-conditioned suppression; MatchDG tests finer correspondence. The ten baselines cover both mechanism controls and DG approaches. MIRO requires a pretrained reference, introducing pretraining-data and model choices beyond our from-scratch protocol.

Unestablished theoretical [...] Reference accuracy is an assumption, not a consequence of neighbourhood tightness (lines 308–310 and 1025). Proposition 2 does not require nuisance-free references: with mi=a(si)+eim_i=a(s_i)+e_i, reference error enters through ei−eje_i-e_j, bounded by 2η2\eta, and a common additive bias cancels. The result characterises approximation conditions and errors; it does not establish that neighbourhood averaging recovers the behavioural component.

Descriptors assess behavioural coherence and alignment ablations assess residualisation; neither directly estimates the unobserved reference error. We have added clarifications (lines 308-310 and 1025).

Inconclusive ablations [...] Full LBNS versus local-uniform tests RBF weighting; local-uniform versus class-uniform compares local and class-wide references. Full LBNS has higher predictive point estimates on both tasks, but paired intervals include zero. These ten-fold controls are exploratory and do not establish component necessity (Appendix A.8, Table 8, lines 957–981).

Table 8 uses identical [...] Equal coefficients in Table 9 need not imply equal regularisation strength. We therefore selected the direct-alignment coefficient by source-validation ROC AUC per fold, keeping the structure coefficient fixed. LBNS has higher mean ROC/PR AUC on five of six datasets and lower domain recoverability on all six (Table 11).

Minor writing issues [...] The introduction now defines behavioural state and nuisance before the equations, with a district heating example distinguishing sensor-calibration offsets from meaningful process changes.

Figure 1 contains [...]. We enlarged Figure 1, added connecting arrows, and explained that local neighbourhoods proxy behavioural similarity, linking the population target to the practical surrogate.

The explanation [...]. Thank you for highlighting this inconsistency. Our wording should have more clearly distinguished the roles of the losses. The structure term encourages compact local groupings among selected same-label neighbours, while the supervised task loss opposes collapse that would erase class discrimination. Residual alignment penalises cross-domain residual mismatch without requiring local behavioural references to coincide. The structure term therefore serves as a local-cohesion regulariser. We have clarified these roles at lines 217 and 223.

Reviewer_4xqaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Existing invariant representation learning methods typically enforce global domain invariance, or condition only on coarse task labels, potentially erasing useful within-label behavioural structure. The work argues and introduces a new, more appropriate target, latent-state-conditioned nuisance invariance (LaSNI), where domain-specific variation is removed only after conditioning on the underlying behavioural state. This target is formally stated as z⊥d∣sz \perp d \mid s, where zz is the learned representation, dd the domain, and ss an unobserved latent behavioural state. Since ss is unobserved, Local Behavioural Nuisance Suppression (LBNS), a source-only surrogate that constructs dynamic same-label neighbourhoods, estimates a local behavioural reference, and suppresses cross-domain variation in the residual rather than the full representation.

Across six real-world tasks and 30 leave-domain-out folds, LBNS achieves the strongest aggregate transfer and lowest average domain recoverability across all eight objectives. LBNS obtains the highest mean ROC and PR AUC on M1, M2, EC, and METR-LA; V-REx is narrowly highest on ELD and MatchDG on BMSAQ. Residual alignment therefore provides a better invariance–fidelity trade-off by preserving locally explained behavioural structure. An independent descriptor diagnostic shows that its latent neighbours are more behaviourally coherent than random same-label samples in every fold and dataset, including under cross-domain restriction.

Critical strengths

Several good practices from the authors.

  1. Explicit claims and assumptions. The authors formulate “latent-state-conditioned nuisance invariance,” state the assumptions motivating it, and also distinguish the proposed target from the practical LBNS surrogate. This allows for scrutiny on the central claim
  2. Very good attention to experimental uncertainty. Reporting results over 30-folds with 95% confidence intervals is a clear strength here, particularly for learning based methods.
  3. Acknowledgement of limits. The authors explicility stated and acknowledge that the proposed neighborhood measure does not establish their original “ideal” objective and the evalution covers only binary tasks and single architecture family.
Critical weaknesses
  1. Mathematical definitions could be stated more explicitly. Definition 1 states “p(z∣d)p(z \mid d) is invariant with respect to d.” Under the usual interpretation, this means p(z∣d)=p(z)p(z \mid d)=p(z), equivalently z⊥dz \perp d, but these different notations forms are used when one could be standard. Consistent presentation of math will help readers understand the ideas behind.
  2. “Nuisance” is a central term in the entire manuscript but it needs a clearer and operational definition. “Nuisance”, “nuisance suppression”, “nuisance variation” cannot be understood clearly without an explicit definition. There are description of nuisance as “domain-specific variations” that remain after “behavioral state” (? clearer definition of this), etc. The authors should explain what makes a component removable or is a nuisance by providing concrete examples.
  3. Lack of explanation on how or why the structure term prevents collapse.
Questions for authors

Please provide clearer and more explicit descriptions of the weaknesses described above and also further clarifications on the following wordings

  1. Section 4, assumption 4 “finer approximation”, finer is what sense? Distance or what measure in what space?
  2. Section 6. “less heterogeneous nuisance to remove”, What heterogeneity and how is the nuisance heterogeneity assessed?
1 comment

We thank the reviewer for the helpful comments.

Mathematical definitions [...] Thank you. We have made Definitions 1 and 2 explicit using independence notation and the corresponding distributional equalities, consistently with Definition 3. This clarifies the notation without changing the definitions or theoretical claims (line 244).

“Nuisance” is a [...] We have added explicit definitions of behavioural state and nuisance variation in the introduction (line 053). Nuisance denotes domain-dependent variation remaining when the underlying process condition is held fixed, such as sensor calibration differences. I.e., nuisance is installation-specific variation that does not reflect meaningful differences in operational behaviour for the task. This distinguishes nuisance from genuine behavioural differences, which we aim to preserve even when informative about domain identity.

The authors should [...] Thank you for this suggestion. We added a concrete industry example at line 056, distinguishing removable sensor calibration offsets from genuine fault-related temperature changes that should be preserved.

Lack of explanation [...]. Thank you for highlighting this point. Our wording should have more clearly distinguished the roles of the losses. The structure term encourages compact local groupings among selected same-label neighbours, while the supervised task loss opposes collapse that would erase class discrimination. Residual alignment penalises cross-domain residual mismatch without requiring local behavioural references to coincide. The structure term therefore serves as a local-cohesion regulariser within the joint supervised objective, rather than independently preventing collapse. We acknowledge this was unclear. We have added clarifications to lines 217 and 223.

Section 4, assumption 4 [...] Thank you. We have replaced "finer approximation" with an explicit expected-distance comparison in the latent behavioural-state space, using the metric ρ\rho defined in Assumption 4 (line 270) and used in Proposition 2. The assumption concerns latent-state proximity, not merely proximity in the learned representation. Because the states are unobserved, the descriptor-based neighbourhood diagnostic provides supporting proxy evidence rather than directly measuring this latent-state distance.

Section 6. [...] Thank you. Here, heterogeneity refers to cross-system differences in equipment, sensing, control, and operating conditions. These differences may involve both genuine behavioural variation and nuisance; we do not independently quantify their nuisance component. We have therefore revised this passage to distinguish the observed performance pattern from its possible explanation. The domain probe measures recoverable domain information, not nuisance heterogeneity itself (line 373).

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

The paper studies domain generalization for heterogeneous physical systems, focusing on the loss of useful within-class behavioural structure caused by coarse invariance objectives. The authors formulate latent-state-conditioned nuisance invariance, which aims to suppress domain-specific variation after conditioning on the underlying behavioural state. The training objective combines residual alignment with supervised classification and local structure regularisation. Experiments on six binary classification tasks show improved aggregate predictive performance and reduced domain recoverability compared with the selected baselines.

Critical strengths
  1. The motivation is easy to follow, and the paper is generally well organized.
  2. Using the same encoder and model-selection protocol across methods helps make the comparisons easier to interpret.
  3. The ablations and domain-probing results offer some insight into how the method behaves beyond its classification performance.
Critical weaknesses
  1. The relevance of several evaluation tasks is unclear. METR-LA is repurposed for morning-evening classification, while ELD labels are constructed from within-day load variability, but the motivation and practical value of these tasks are not sufficiently explained. The tasks of these datasets are also not the typical tasks for these datasets. The paper also provides limited details on their construction and does not compare against simple task-specific rules, making their difficulty hard to assess. Evaluation on established benchmarks would be needed to provide more convincing evidence.

  2. Baseline coverage is limited. Representative DG methods are missing. Although the selected comparisons help examine different invariance objectives, they do not sufficiently establish how competitive LBNS is against the broader DG literature.

  3. The main motivation closely follows MatchDG, which already identifies limitations of class-conditional invariance and learns finer-grained cross-domain matches when the underlying objects are unobserved. The paper needs to demonstrate more clearly what substantial new insight its continuous-state formulation provides beyond this existing framework.

  4. The hyperparameter search is not described in enough detail to assess the fairness of the comparison. LBNS uses dataset-specific settings for the neighbourhood size, kernel bandwidth, and both regularisation coefficients, whereas baselines use fixed settings across datasets.

Questions for authors
  1. How are the METR-LA input windows and ELD variability labels defined, and what motivates these classification tasks?
  2. How well do simple statistical or rule-based classifiers perform on the constructed METR-LA and ELD tasks?
  3. Is the improvement offered by the proposed method significant compared to the strongest baseline?
1 comment

We thank the reviewer for the helpful comments.

How are the [...] Thank you for highlighting this ambiguity. METR-LA uses full-day profiles of 288 five-minute readings, classified by evening-minus-morning contrast (mean 17:00–21:55 minus mean 06:00–09:55), with low/high labels below/above each sensor’s median contrast. ELD uses 96 fifteen-minute readings daily, with low/high variability defined by daily standard deviations below each client’s 20th or above its 80th percentile, discarding intermediate days.

These tasks complement field fault detection with more standardised measurements. Their coarse, domain-relative labels group different daily profiles, supporting tests of conditioning granularity. Section 5 now clarifies their construction, sample selection and rationale.

How well do simple [...] Across the same 30 held-out-domain folds, a linear-trend threshold, nearest centroid, Gaussian Naïve Bayes, and decision tree classifiers achieved ROC/PR AUCs of 63.86%/64.55%, 69.68%/71.36%, 52.45%/57.17%, and 67.14%/65.74% on ELD, and 74.68%/68.68%, 85.39%/85.87%, 85.30%/86.32%, and 83.00%/79.87% on METR-LA, respectively.

However, our objective is not to establish that a Transformer is the best classifier. We aim to isolate how nuisance-suppression objectives affect transfer and domain recoverability. Simple classifiers characterise task difficulty; these datasets complement the suppression analysis without implying that sophisticated modelling is necessary.

Baseline coverage is limited. [...]. We added SWAD (weight averaging), CCIL (class-wise feature-contribution consistency), and FIXED (feature mixup, adversarial domain learning and large-margin discrimination) across all six datasets and the same 30 held-out-domain folds.

These complement ERM’s no-suppression reference, global suppression (DANN), class-conditioned mean/adversarial alignment (CCA/CDANN), fine-grained correspondence (MatchDG), latent-subdomain learning (DIVERSIFY), and risk invariance (V-REx). The ten comparators cover both conditioning granularity and alternative DG mechanisms. A shared time series encoder helps isolate differences between learning objectives. Our aim is to test conditioning granularity and the invariance–fidelity trade-off, with conclusions scoped to the evaluated settings.

Is the improvement [...] We added paired significance tests and effect sizes (Appendix A.5, Table 5; Results, line 408).

LBNS has the highest aggregate ROC/PR AUC (76.49%/74.48%, versus 73.34%/70.18% for CCA, the next-best aggregate baseline) and lower chance-normalised domain recoverability (2.071 versus 3.478). Exploratory pooled paired tests over repeated, overlapping folds yield p<0.001p<0.001 for all three metrics, with dz=0.84d_z=0.84, 0.990.99 and −2.16-2.16 for ROC, PR and domain recoverability, respectively.

LBNS also has the best average fold rank (3.84; MatchDG: 5.01; CCA: 5.21). Dataset-level comparisons show significant M1 ROC and EC PR gains against both comparators, and significantly lower domain recoverability in nine of twelve comparisons. The main finding is strong aggregate transfer alongside reduced recoverable domain information, with predictive gains on some tasks and competitive performance on others. This supports our motivation for suppressing nuisance while preserving behavioural information.

The main motivation [...] We agree that MatchDG already identifies limitations of class-conditional invariance and motivates finer-grained correspondence. Our distinction concerns what should be aligned when comparable behavioural states are potentially non-identical. MatchDG encourages agreement across inferred object matches, whereas LBNS aligns residuals around sample-specific local references. Its suppression term therefore permits zi−zj≈mi−mjz_i-z_j \approx m_i-m_j, allowing locally explained differences between nearby operating states to remain. This matters for continuously varying load, operating regime and fault severity, where behavioural comparability need not imply identical representations. We test this distinction through the matched direct-versus-residual alignment control, including source-validation selection of the direct-alignment coefficient. We have clarified this contribution in Related Work.

The hyperparameter [...] Thank you for identifying this issue; we agree this was missing. All methods selected hyperparameters using source-validation performance only, with method-specific search spaces and no held-out target performance. DIVERSIFY and V-REx initially used a pooled source-validation search over M1, M2, and EC. We have now repeated these searches and selected configurations independently for each dataset to match the other methods. We clarify these selection procedures in Appendix A.4 and have updated the manuscript accordingly.

Comments on the paper

2 comments

Dear Reviewers and Area Chair,

Thank you for your constructive feedback. We have uploaded a revised manuscript, with changes highlighted in red. Our point-by-point responses explain the clarifications and additional experiments and indicate the relevant sections of the revised manuscript.

Best regards, The authors

Dear Reviewers and Area Chair,

As the author-reviewer discussion period draws to a close, we would like to thank you again for your feedback.

We hope that our point-by-point responses, additional experiments, and revised manuscript have addressed the concerns raised. To date, we have not received further questions; should any concerns remain unresolved, we would appreciate the opportunity to clarify them before the discussion period closes.

Thank you again for your time and consideration.

Best regards, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.