acceptodds
Under review as a conference paper at ICLR 2027

Sequential Data Assimilation using Rolling Diffusion Models with Editable History

Abstract

Data assimilation (DA) combines dynamical models with observations to estimate the hidden states of complex systems. Classical methods such as the ensemble Kalman filter struggle under highly non-Gaussian dynamics, while existing diffusion-based approaches are either fixed-window smoothers or bias-prone online filters. We introduce Rolling Diffusion Data Assimilation (RDA), which uses rolling diffusion models for forecasting, and assimilating data sequentially by guiding the denoising process with observations. Additionally, we propose a new rolling diffusion architecture with editable history blocks that can revise past frames as new observations arrive. For an idealized version with exact conditional denoising, our analysis identifies the limiting window posterior and bounds its discrepancy from the full filtering distribution under strong mixing. Across a range of nonlinear dynamical systems, from low-dimensional chaotic dynamics to high-dimensional geophysical settings, RDA matches or outperforms both classical and ML-based filtering methods, especially under sparse and nonlinear observations.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Data assimilation, rolling diffusion, filtering, forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_opW4OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This submission addresses the data assimilation problem for weather forecasting. The authors propose Rolling Diffusion Data Assimilation (RDA), a new method based on rolling diffusion models that denoises a sequence of states with sequential noise levels. The authors argue that current methods fail to properly account for the uncertainty of past states and propose addressing this via rolling diffusion: a sampling algorithm that regenerates past states based on their posterior distribution. The method is evaluated on Lorenz 63, surface quasigeostrophic dynamics, and ERA5, demonstrating superior posterior sampling abilities over the baselines. The authors also present several ablations to assess the effect of different hyperparameters.

Critical strengths

The method is evaluated on several dynamical systems of increasing complexity. The Lorenz 63 evaluation helps visualize and understand the method, the ERA5 dataset serves as the targeted real-world application, and the surface quasigeostrophic dynamics model offers appropriate intermediate complexity.

The experiments demonstrate that the proposed method improves performance over all compared baselines when evaluated on suitable probabilistic metrics.

Critical weaknesses

The exposition of the method is fairly complex and difficult to follow overall due to heavy notation. Furthermore, the paper provides little intuition regarding the fundamental limitation identified by the authors and how the proposed solution resolves it.

1. Novelty

1.1 The method's novelty is not adequately discussed relative to recent diffusion-based and diffusion-forcing methods. In the methodology section, the authors first present rolling diffusion, an established technique, and then present a straightforward extension to account for observations in Section 3.2. As presented, the paper's primary contribution relies on the editable history framework presented in Section 3.2 to bridge a theoretical gap. To convincingly demonstrate the impact of this technique over a simpler adaptation of rolling diffusion, the effect of editing history should be explicitly evaluated and discussed. However, this comparison is barely studied in the experiments.

1.2 The core limitation identified by the authors is also not clearly illustrated. If this theoretical gap is indeed a primary drawback of prior work, a concrete example should demonstrate what a data assimilation method fails to capture when it does not model the smoothing distribution. Instead, the paper only reports benchmark tables with aggregated metrics.

1.3 In the introduction, the authors highlight the high memory and inference costs of score-based data assimilation and claim that RDA addresses this issue. However, it remains unclear from the text how RDA achieves these computational savings.

2. Empirical validation

2.1 I have concerns regarding the choice of baselines. In both the surface quasigeostrophic system and ERA5 experiments, RDA is compared only against another deep generative model. It should be benchmarked against operational state-of-the-art baselines like 4D-Var or the Ensemble Kalman Filter (EnKF). Without these standard baselines, the evaluation relies entirely on neural-network-based methods and their specific training setups, making the claims less convincing.

2.2 Furthermore, while the authors discuss theoretical differences between RDA and other diffusion-based data assimilation methods, these alternative diffusion baselines are not evaluated on the ERA5 dataset. Consequently, RDA is compared against only a single deep learning baseline on ERA5.

2.3 Finally, the motivation for using quadratic or tanh observation operators is unclear. The physical motivation for these observation processes should be explicitly justified.

3. Relevance of the theoretical guarantee

The theoretical analysis assumes exact Bayesian conditional updates (Theorem 2). However, diffusion posterior sampling algorithms provably yield a biased posterior, even in the Gaussian case [1]. Consequently, the applicability and usefulness of this theoretical result are unclear, providing little practical intuition for the proposed method.

References

[1] Hodyss, D., & Morzfeld, M. (2026). Using diffusion models to do data assimilation. Monthly Weather Review, 154(2), 165-182.

Questions for authors

1. Since the history of past states is editable and variable in the proposed framework, how is temporal consistency between consecutive reconstructed states ensured?

2. What is the underlying backbone generative architecture, and how is it trained?

3. What is the proposed method's computational cost in terms of memory consumption and the number of neural network evaluations, and how does it compare quantitatively to previous generative data assimilation methods?

Reviewer_K6X7OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The aim of this paper is to propose a Bayesian filtering method that estimates the current state of a complex dynamical system from a sequence of observations. This amounts to sampling from the filtering distribution p(xn∣y1:n)p(x^n | y^{1:n}). To this end, the authors propose a method based on rolling diffusion models: observations are incorporated sequentially using posterior sampling methods and previous states can also be corrected by renoising them at a chosen noise level and then denoising them using the new observations. The proposed method achieves strong results across all experiments, including Lorenz-63, SQG, and ERA5.

Critical strengths
  • The paper is well written and well organized.
  • The related work section is comprehensive, clear, and concise.
  • The filtering problem is of great practical importance, as it is addressed daily by operational weather forecasting centers.
  • Experimental results outperform existing state-of-the-art methods.
Critical weaknesses
  • The proposed editing strategy reminds me of the renoising procedure used in DAISI [1]. Both approaches perturb clean states by injecting a controlled amount of noise (determined by the noise schedule σ0(1)\sigma_0(1) in this paper and by tmin⁡t_{\min} in DAISI) and then incorporate observations through training-free posterior sampling methods. However, the ability to correct the state depends on the noise level: more noise allows greater corrections and better integration of the observation, but also risks destroying the existing structure. In DAISI, this trade-off is controlled by the hyperparameter tmin⁡t_{\min}, which has a non-negligible impact on the discrepancy between the resulting distribution and the true filtering distribution. In this paper, the amount of injected noise is fixed during training through σ0(1)\sigma_0(1) and I would therefore be interested in understanding how sensitive the results are to this parameter.

  • As stated in Equation (8), previous states of the dynamical system must be sampled from the smoothing distribution p(x0:n−1∣y1:n)p(x^{0:n-1} | y^{1:n}). However, it is unclear to me how applying the proposed sequential method yields samples from this distribution. My current intuition is that the samples are first denoised using their corresponding observations and then edited through renoising to incorporate information from "future" observations. However, I believe the paper would benefit from a more detailed explanation of this.

  • The method's performance depends on the total window length LL, and training a generative model with a large LL may be hard for high-resolution atmospheric systems. Inference can also be slow, as methods such as MMPS require computing a JVP at each diffusion step, and may also introduce practical biases, particularly when the observation operator is nonlinear. However, these limitations are shared with other methods, and I do not consider them as major concerns, but rather practical aspects worth keeping in mind for operational deployment.

  • In the first two experiments, the systems considered are deterministic. Consequently, particle-filter-based methods (FlowDAS and FA-APF) reduce to standard bootstrap particle filters (since the optimal and standard proposals coincide in this setting). Thus, including these methods in these settings may be redundant: the authors could instead consider using a standard bootstrap particle filter (BPF) as a baseline or evaluating the methods on stochastic systems.

[1] DAISI: Data Assimilation with Inverse Sampling using Stochastic Interpolants, Andrae et al., ICML 2026.

Questions for authors

Could the authors please clarify the first two points raised in the weaknesses? I would be happy to increase my score if these concerns are adequately addressed, or if the authors can explain why my reasoning is incorrect.

Reviewer_oxa5OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Rolling Diffusion Data Assimilation (RDA), an online filtering method that combines rolling diffusion with observation-guided denoising. To address the limitations of conditioning on fixed historical states, RDA introduces editable history frames that are refined using recent observations. The authors show that, under idealized conditions, repeated refinement converges to the posterior given recent observations, which approaches the true filtering distribution as more history is retained. Experiments on Lorenz–63 and SQG show improved filtering accuracy and lower computational costs than several existing methods. On ERA5, RDA achieves results close to those of an offline smoother.

Critical strengths

Clear motivation. Equation (8) identifies a limitation of forecast-native generative filters that condition on fixed historical states. This provides a direct motivation for introducing editable history.

ForcingDAS-Rolling improves over ForcingDAS-AR across all evaluated Lorenz–63 and SQG settings without retraining. This suggests that the proposed inference schedule can benefit different generative priors.

The paper includes classical and generative filtering baselines evaluated using shared observations. The comparison against a 10510^5-particle bootstrap filter on Lorenz–63 is useful for assessing posterior quality rather than only reconstruction accuracy.

Reusing partially denoised windows reduces repeated sampling. The reported SQG experiments show 3.9–5.6× speedups over DAISI and 37–53× over FlowDAS.

Appendix D.6 explicitly acknowledges where the theoretical assumptions differ from the practical implementation, helping readers interpret the guarantees.

Critical weaknesses
  1. The paper has limited methodological novelty. ForcingDAS already supports flexible noise schedules and the refinement of past states through fixed-lag smoothing. The paper also shows that RDA's inference schedule improves ForcingDAS without retraining, suggesting that much of the benefit comes from the scheduling strategy rather than a new modeling approach. However, it remains unclear what RDA adds beyond these existing capabilities and whether its specially trained prior provides additional benefits.The authors should clarify the distinction from ForcingDAS and compare against its fixed-lag smoothing approach adapted for online filtering.

  2. The theory assumes exact sampling and infinitely many refinement passes (K→∞K\to\infty), while the experiments use only one pass (K=1K=1) with approximate guidance. The theoretical limit also discards information carried forward from previous assimilation steps, which may be important to RDA's performance. The experiments highlight this gap. On Lorenz–63, increasing KK improves accuracy but never matches the result obtained using a single pass with tuned guidance. This suggests that guidance tuning and retained history may contribute more to the improvements than the theoretical mechanism. Moreover, the strong-mixing assumptions do not hold for the deterministic Lorenz–63 and SQG systems. The authors should clarify how the theory relates to the implemented method or validate it on a system where the assumptions hold.

  3. The paper allows past states to be revised as new observations arrive, but does not show how much these states actually change. Since history frames are re-noised to relatively low noise levels (only to σ0(1)\sigma_0(1)), it is unclear whether the method can make substantial corrections or recover different posterior modes. Furthermore, the history-length ablation does not establish whether improvements come from editing past states, using longer temporal context, or incorporating past observations. The authors should measure how much history changes and compare the full method against versions where historical states remain fixed or are guided only by the current observation.

  4. Experimental limitations.

    a. Unclear source of performance gains. RDA outperforms ForcingDAS-Rolling, but the two use different architectures and training procedures. It is therefore unclear whether the improvement comes from the proposed method or the different learned priors. A more controlled comparison is needed.

    b. Unclear reliability of posterior estimates. RDA uses strong observation guidance, which may improve reconstruction accuracy while reducing ensemble diversity. The paper does not directly evaluate whether the resulting uncertainty estimates are well calibrated. Additionally, tuning on a single Lorenz–63 validation trajectory raises questions about robustness.

  5. Limited discussion of related flow-matching methods. The paper could better position RDA relative to recent flow-matching approaches. EnFF (Transue et al.) provides a training-free filtering alternative but is not evaluated. Accelerated Sequential Flow Matching (Huang et al., 2026) also uses the previous posterior to accelerate sequential sampling, while FlowDA (Cheng & Zhu, 2026) addresses efficient weather-scale assimilation. Although these methods target different settings, discussing their connections to RDA would help clarify its contribution. An EnFF comparison on a shared benchmark would also be useful if feasible.

Transue, T., Chen, B., Takao, S., & Wang, B. (2025). Flow Matching for Efficient and Scalable Data Assimilation. https://arxiv.org/abs/2508.13313

Huang, Y., Hsu, H. H.-H., Wang, J., Dai, B., & Li, P. (2026). Accelerated Sequential Flow Matching: A Bayesian Filtering Perspective. https://arxiv.org/abs/2602.05319

Cheng, R., & Zhu, L. (2026). FlowDA: Accurate, Low-Latency Weather Data Assimilation via Flow Matching. https://arxiv.org/abs/2602.06800

Questions for authors

When changing the history length HH, is the diffusion prior retrained, or is only the inference configuration changed? Also, how is the initialization noise schedule used with βtrain=0.5\beta_{\mathrm{train}}=0.5 constructed?

Why does RDA require large guidance strengths in several experiments? How does the optimal guidance strength change with history length and the amount of re-noising applied to past states?

The paper reports substantial runtime improvements, but how do the number of denoiser evaluations (NFEs) and FLOPs per assimilation step compare across RDA, SDA, and DAISI? This would help clarify the source of the efficiency gains.

Algorithm 2 describes re-noising for variance-exploding (VE) diffusion. How is the same operation implemented for the variance-preserving (VP) diffusion models used in the experiments?

Since ERA5 assimilation operates in latent space, how are physical-space observations incorporated? Is the latent state decoded to compute the observation likelihood, and does MMPS differentiate through the decoder?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.