acceptodds
Withdrawn from ICLR 2027

A Hierarchical Bayesian Approach for Non-Markovian Evolving Domain Generalization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Evolving Domain Generalization (EDG) considers learning from a sequence of domains whose underlying distributions change over time, reflecting the non-stationary nature of many real-world learning environments. Recent structure-aware methods model this evolution by separating static task factors from dynamic domain factors that evolve across domains. However, their autoregressive latents impose a first-order Markov assumption, tying the current domain state to its immediate predecessor, which is restrictive when the dynamics have persistent temporal structure or abrupt transitions. We propose a hierarchical state-space model (SSM) formulation that introduces regime latents governed by change-point variables. Moreover, the formulation naturally casts EDG as a sequential Bayesian inference problem, which we leverage Rao-Blackwellized particle filter (RBPF) to approximate the posterior over the latents. Experiments on synthetic benchmarks containing both gradual and abrupt distribution shifts and real-world evolving domain benchmarks show that the proposed approach improves mean target-domain accuracy by 17.06 points over the strongest baseline on the original streams and by 15.74 points on the abrupt variants. An ablation shows that the change-point latent contributes 12.47 points of abrupt-stream accuracy. These results highlight the value of hierarchical sequential Bayesian inference for modeling complex dynamics in evolving domains.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Evolving domain generalization, state-space model, sequential Monte Carlo, Rao-Blackwellized particle filter, change-point detection

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_7YRgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies non-Markovian evolving domain generalisation (EDG), where domains drift continuously within a regime but the drift mechanism itself can change abruptly. The authors propose a hierarchical switching state-space model in which a change indicator ct and a regime variable rt govern the transition statistics of a continuous domain state zt, and a Rao-Blackwellized particle filter (RBPF) jointly infers the discrete regime history and the continuous state. The method includes a hazard process, regime priors, an inner moment-matching update, particle weights, change-point readout, and resampling. The authors also construct abrupt EDG benchmarks (Circle, Sine, RMNIST, PowerSupply, Portraits, CalTran) with controlled shifts. Table 2 reports mean target accuracies of 88.63% on the original streams and 82.58% on the abrupt variants. The paper concludes that explicit regime and change-point inference improves post-change generalisation over a continuous-only SSM.

Critical strengths

Separating within-regime gradual drift from a change in the drift mechanism is well motivated, and the hierarchy in Section 3.2 gives the two effects separate, interpretable roles; this is more suitable for abrupt EDG than a single continuous random walk.

Section 3.3 gives a concrete RBPF procedure, including the inner moment-matching update, particle weights, change-point readout, and resampling; the authors are also candid about the limits of change-point identification (Section 5, Figure 5).

Critical weaknesses
  1. I had difficulty reconciling the filtering procedure with the EDG setting. Section 3.1 (p. 4) defines each domain as a labeled batch. Eq. (5) (p. 5) uses its labels in the likelihood, and Algorithm 1 (p. 6, steps 8-11) updates the posterior before predicting labels. The loop covers the full stream without a target-domain branch. Read literally, predictions can depend on labels from the domain being evaluated.

Section 4.1 (p. 7) rules out target labels for hyperparameter selection, but does not settle their use in filtering. Section 5 (p. 9) says predictions precede the next domain's observations. If the reported scores are one-step-ahead predictions, this indexing needs to be explicit in the algorithm and prediction formula. If there is instead a labeled support set, its separation from the scored examples and the baselines' label access need to be stated.

Eq. (5) also cannot by itself explain an unlabeled target update: marginalizing its unknown labels gives one. I cannot establish label leakage without the code, but the information used for the scored predictions is not clear enough to interpret Table 2.

  1. Table 3 (p. 8) matches the observation family and particle count; Appendix C (p. 14) lists D=4, K=600, and L=8 for both Circle/Sine methods. However, Section 3.3 (p. 6) describes the continuous-only control as a bootstrap filter, while the full method uses Gaussian moment matching. As written, the comparison changes both the regime variables and the treatment of the continuous posterior. Keeping the same inner update and fixing the regime would isolate switching more clearly.

The gains are uneven. Abrupt Circle rises from 43.70% to 95.90%, while Sine and PowerSupply gain only 0.88 and 0.57 points. Section 4.2 (p. 8) attributes the Circle result to the invert bit, also specified in Appendix B (p. 13). This supports the full mechanism on the constructed reversal, but does not isolate how much of the average 12.47-point gain comes from the change-point indicator itself.

  1. In Figure 5 (p. 9), the first dashed change marker appears at t=8 for both Sine and RMNIST. Table 1 (p. 7) and Appendix A (p. 13) specify t=12 and t=10, respectively. The second markers agree at t=16 and t=13, so a zero-based versus one-based indexing difference does not explain this. Were the curves generated with the same configurations as the accuracy tables? This needs resolving before the claim of detection without delay can be assessed.

The RMNIST curve also has additional high-probability alarms, as the caption acknowledges. Section 4.3 gives no detection-delay or false-alarm statistics, leaving the reliability of the detector difficult to judge.

Questions for authors

Does Table 2 score current-domain predictions after updating with Dt labels, or next-domain predictions before Dt+1 is observed? Which labels enter Eq. (5) and Algorithm 1? Please give the exact evaluation order, any support/query split, and the baselines’ label access.

Does the continuous-only control use the same Gaussian moment-matching inner update as the full method? If not, can you provide a matched control with fixed regime and the same inner update to isolate switching? Please also report variance or independent runs for the sub-percentage gains on Sine and PowerSupply.

Which first-change locations were used for Figure 5? Please reconcile t=8 with Table 1’s t=12 and t=10, and report detection delay and false alarms. Also clarify why Circle’s first change at t=10 lies within the 15/5/10 source prefix, while Section 4.1 says jumps are placed after the source split.

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

This paper proposes a hierarchical switching state-space model (SSSM-PF) for evolving domain generalization, aiming to handle both gradual distribution shifts and abrupt changes in domain evolution. The method introduces discrete change-point and regime variables alongside continuous domain states, and employs Rao-Blackwellized particle filtering for sequential Bayesian inference.

Experiments on six original EDG benchmarks and their abrupt-shift variants report improvements of 17.06 and 15.74 percentage points, respectively, over the strongest baselines in the corresponding settings. Ablation studies further investigate the contribution of explicit change-point modeling.

Critical strengths

Explicit modeling of regime changes. The paper addresses a meaningful limitation of EDG methods that assume relatively homogeneous temporal dynamics. Separating gradual within-regime evolution from abrupt changes in the underlying transition mechanism is a reasonable modeling direction.

Evaluation under abrupt distribution shifts. The construction of abrupt-shift benchmarks and the change-point ablation provide useful experimental settings for investigating regime-aware domain evolution. The reported improvements are substantial, particularly on synthetic benchmarks.

Critical weaknesses

Potential target-label leakage in the evaluation protocol. This is my primary concern. Section 3.3 and Algorithm 1 describe filtering updates based on the likelihood (p(Dt∣zt)p(D_t \mid z_t)), where (DtD_t) contains labeled samples. The observation model in Equation (5) explicitly depends on the labels. Moreover, Algorithm 1 performs the likelihood-based update before producing predictions.

As written, this suggests that predictions for a target domain may be generated using a posterior that has already incorporated labels from that domain. If this is how the evaluation is implemented, it would violate the standard EDG protocol and fundamentally undermine the comparison with baselines that do not access target labels.

The statement that target labels are not used for hyperparameter selection does not resolve this issue, since posterior inference and hyperparameter selection are different operations. The paper must clearly establish that target-domain predictions are made without incorporating unavailable target labels.

Heavy reliance on task-specific observation models weakens the generality of the contribution. Appendix B specifies substantially different latent representations and observation models for different benchmarks, including polynomial decision boundaries for Circle/Sine, inverse geometric transformations for Rotated MNIST, and time-indexed calibration for Portraits/CalTran. These choices introduce substantial structural knowledge about the underlying shifts.

Consequently, the large performance gains may partly reflect favorable task-specific inductive biases rather than the effectiveness of hierarchical Bayesian inference itself. This is particularly concerning when comparing against general-purpose EDG baselines. Moreover, the claimed non-Markovian modeling contribution should be better distinguished from standard switching state-space models, which remain Markovian in their augmented latent state.

Questions for authors

Clarification of the evaluation protocol: Are target-domain labels ever used in the likelihood updates of Algorithm 1 before computing target-domain accuracy? Please explicitly describe the prediction and filtering-update order for source, intermediate, and target domains. If the current evaluation incorporates target labels, can the authors report results under a strictly label-free target evaluation protocol?

Clarification of the structural assumptions: How much of the improvement is attributable to the task-specific observation models and shift-type priors? Can the authors demonstrate the benefit of the switching inference mechanism under a common observation model or when the true shift mechanisms are not specified in advance?

Reviewer_Voy3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies evolving domain generalization under gradual distribution shifts and abrupt changes in domain evolution mechanisms. It proposes a hierarchical switching state-space model that introduces regime latents governed by change-point variables. The model jointly represents continuous within-regime evolution and abrupt regime transitions without specifying the number of regimes in advance. A Rao–Blackwellized particle filter approximates the posterior over continuous domain states and regime changes for prediction and change-point detection. Experiments on original evolving-domain benchmarks and manually constructed abrupt variants report improved target-domain accuracy over the compared baselines.

Critical strengths
  1. The paper addresses a practical and important problem.

  2. The hierarchical formulation explicitly separates within-regime evolution from regime changes, providing a clear modeling perspective.

  3. The Bayesian framework maintains uncertainty over domain states and regime changes and produces an explicit change-point probability.

  4. The authors provide implementation code and dataset-generation scripts, facilitating reproducibility and inspection.

Critical weaknesses
  1. It is unclear how the model identifies an unseen label reversal without access to target labels. In abrupt Circle, the labeling rule reverses at domain 20, exactly where the target split begins, while neither the source nor intermediate domains contain this reversal. The same geometric evolution could occur with or without reversing the labels, so the observed inputs alone do not determine which labeling rule applies. The paper does not explain what observable evidence enables it to select that hypothesis and achieve 95.90% target accuracy.

  2. In the introductory example, illumination and weather can change at the same time, so it is unclear why they should be treated as separate stages. Why not model their joint effects directly, instead of organizing these changes around time? The paper needs to explain what regime switching adds beyond modeling the underlying factors.

  3. The method uses polynomial decision boundaries, inverse geometric transformations, and temporal calibration for different datasets. The experiments do not adequately establish whether the gains come from the proposed hierarchy or these task-specific assumptions.

  4. The distribution shifts in PowerSupply and CalTran arise from the reordering of data segments. They do not necessarily imply the emergence of new generative mechanisms, as might occur in real-world environments. The current datasets and baselines are still relatively weak.

Questions for authors
  1. How does the model identify regime changes in target domains without accessing target labels?

  2. How does the proposed problem setting differ from temporal domain generalization? e.g., [1, 2, 3].

  3. Why use temporal regimes rather than directly model jointly varying factors such as illumination and weather? If the jointly varying factors driving distribution shifts are modeled directly, is an additional regime variable still necessary?

[1] Generalizing across temporal domains with koopman operators, AAAI24 [2] Continuous temporal domain generalization, NeurIPS24 [3] Learning robust spectral dynamics for temporal domain generalization, NeurIPS25

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

Related papers

Discussion (0)

Sign in to comment.