acceptodds
Under review as a conference paper at ICLR 2027

DAWIS: Data Assimilation with Windowed Inverse Sampling via Multitask Interpolants

23 views

Abstract

Flow- and diffusion-based generative models have recently emerged as flexible and highly efficient forecasting models for dynamical systems. When combined with inference-time guidance, they offer a promising route to high-dimensional non-Gaussian data assimilation (DA), the problem of combining forecasts with observations to estimate latent system states. Existing filters, however, condition on a fixed history and assimilate only the most recent observation, leaving them unable to revise past states when new observations arrive. Estimates then stay tethered to a history that later observations may contradict, and errors accumulate over the assimilation run. To this end, we introduce DAWIS, a unified DA method covering filtering, fixed-lag smoothing, and block smoothing within a single framework. DAWIS replaces the single flow time of a state-level prior with a multitask stochastic interpolant over a window of consecutive states, assigning a separate flow time to each. An assimilation cycle inverts the window to a vector of per-state turning points and regenerates it under observation guidance, with the turning points controlling how strongly each state is held fixed, revised, or generated from scratch. The same construction can also absorb the forecast into the assimilation cycle, removing the need for a separate forecasting model. Experiments on challenging nonlinear systems show that DAWIS improves on both filtering and smoothing baselines under sparse, noisy, and nonlinear observations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Data assimilation, Generative modeling, Filtering, Inverse problem, Stochastic Interpolants, Weather, Conditional sampling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_YwYgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DAWIS, a generative data assimilation framework that partially inverts and regenerates windows of states using multitask stochastic interpolants. Per-state flow times enable filtering, fixed-lag smoothing, and block smoothing with a shared prior, while DAWIS-Joint eliminates the external forecast model. Experiments on SQG and SEVIR show promising improvements over existing filtering and smoothing methods.

Critical strengths

The windowed inversion mechanism addresses error accumulation by allowing new observations to revise previous state estimates. Filtering and fixed-lag smoothing are obtained from the same assimilation cycle.

The framework provides a useful unification of existing generative DA approaches, including forecast-free assimilation and block smoothing.

The experiments cover challenging nonlinear and multimodal observation settings, with comprehensive evaluations of accuracy, calibration, and spectral fidelity. The SQG improvements are particularly convincing.

Critical weaknesses

Unclear posterior correctness. The filtering and fixed-lag smoothing procedures lack theoretical consistency guarantees. Previously assimilated states are partially inverted and guided again on observations already incorporated in earlier cycles. It is unclear whether this repeated conditioning introduces posterior bias or overconfidence. Proposition 1 establishes invariance only for the block smoother under exact guidance. The implementation, however, uses approximate MMPS guidance and sometimes large guidance strengths (ζ up to 20 for filtering and 150 for block smoothing). A comparison against exact Kalman/RTS posteriors on a linear-Gaussian system would directly address this concern.

Incremental methodological novelty. DAWIS combines inverse sampling from DAISI [1], multitask interpolants [3], and per-state noise schedules [4], which ForcingDAS [2] already uses for unified DA. The block-smoothing invariance result is closely related to concurrent work [5, 6]. DAWIS differs from ForcingDAS in revising the anchor state and in smoothing full trajectories without a full-sequence prior. Still, the main contribution is an effective combination of existing techniques. The benefit of revising previously fixed states is also not isolated, for example by an ablation with τ_min = 1 on all retained slots.

Limited ablations and statistical evidence. Sensitivity to window length, turning-point schedules and guidance strength is not systematically evaluated. Results use only 10 trajectories, and the large SEVIR standard deviations make several improvements difficult to assess; paired per-trajectory comparisons would help. The practical relevance of the block-smoothing guarantee is also unclear, since the Soft Block Smoother, which has no such guarantee, performs comparably or better (Table 17).

References

  1. Andrae, M., Wikingsson, E., Takao, S., Landelius, T., & Lindsten, F. (2026). DAISI: Data Assimilation with Inverse Sampling using Stochastic Interpolants. ICML 2026.
  2. Jia, Y., Chen, S., Pan, Y., et al. (2026). ForcingDAS: Unified and Robust Data Assimilation via Diffusion Forcing. arXiv:2605.14285.
  3. Negrel, H., Coeurdoux, F., Albergo, M. S., & Vanden-Eijnden, E. (2025). Multitask Learning with Stochastic Interpolants. NeurIPS 2025.
  4. Chen, B., Martí Monsó, D., Du, Y., Simchowitz, M., Tedrake, R., & Sitzmann, V. (2024). Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion. NeurIPS 2024.
  5. Hill, M. (2026). Exact Global MCMC with Denoising Diffusion. arXiv:2609.00279.
  6. Kang, H., Levi, N. I., Wegner, C. E., Korchinski, D. J., & Wyart, M. (2026). Sampling Data with Chains of Forward-Backward Diffusion Steps. arXiv:2605.27006.
Questions for authors
  1. Can the authors characterize the distribution targeted by sequential windowed inversion, and explain whether repeated observation guidance introduces posterior bias? A linear-Gaussian comparison against exact Kalman/RTS posteriors would help.

  2. What improvement comes specifically from revising the previously fixed historical state? Since Joint AR and ForcingDAS-Pyr already share the DAWIS prior, can this be isolated at matched smoothing lags and computational budgets? Currently their effective lag is 5 versus 6 for DAWIS, and ForcingDAS-Pyr uses 20 solver steps per cycle versus 100.

  3. How sensitive are the results to window length, turning-point profiles, and guidance strength? For the filter and lagged smoother, would direct SDEdit-style noising, as already used for the block smoother (Eq. 9), perform comparably to stochastic inversion?

  4. Can the authors report inference costs and paired statistical comparisons, particularly on SEVIR? Also, Section 5 reports 384×384 resolution, whereas Appendix D.5 specifies 128×128; which was used?

2 comments

We thank the reviewer for their time and detailed review. We appreciate the positive comments about the usefulness of DAWIS and hope to clear out any concerns below. We refer to two common responses: A (the DAWIS target) and B (new experiments).

[W1], [Q1]: Unclear posterior correctness.

We have characterized the DAWIS target in Appendix D.1. The final expression and the interpretation of double counting observations can be found in common response A.

[W1], [Q1]: Linear-Gaussian Comparison.

We thank the reviewer for this suggestion and have performed an experiment on a linear-Gaussian system to illustrate the bias and hyperparameters (A3).

[W1]: Biased guidance methods.

While the paper uses MMPS, this is just one possible choice, and better guidance methods are directly applicable to DAWIS. For Saturating we used ζ=20\zeta=20 for the filter and ζ=150\zeta=150 for the block smoothers. With arctan observations, the observation Jacobian is nearly zero in the saturated region, so the linearised MMPS likelihood gradient is far too weak and the strength compensates for this.

[W2]: Incremental methodological novelty.

Kang et al. and Hill et al. are both cited in the paper and should be viewed as concurrent work. Further, these papers concern only the noise-denoise update and not its application to smoothing posteriors which introduces additional challenges such as cycling. To the best of our knowledge, the use of windowed flow priors for block smoothing, including both the Blocked Gibbs Smoother and the DAWIS Block Smoother, has not been considered before.

[Q2]: Matched inference-costs to ForcingDAS-Pyr.

We rerun ForcingDAS-Pyr with 200 solver steps per cycle, which costs more than a DAWIS cycle. A larger budget does not improve ForcingDAS-Pyr relative to DAWIS in any setting (B7).

[Q2]: ForcingDAS-Pyr and Joint AR at Matched Lag.

ForcingDAS-Pyr and Joint AR with W∣1W\vert 1 have an effective smoothing lag of 5, compared with 6 for DAWIS. We additionally evaluate DAWIS and DAWIS-Joint at an effective lag of 5 (B6). Both DAWIS methods retain stronger performance than the corresponding baselines at this matched lag.

[W3], [Q3]: Window size ablation.

We have performed an ablation on the window size w=2,4,6w=2,4,6 (B1). Throughout all experiments, DAWIS is the least affected by ww, and its w=2w=2-variant still outperforms all other filters using w=6w=6. We can also clearly see the degradation of Joint AR W∣1W\vert 1 at shorter windows.

[W3], [Q3]: Turning point ablation.

In common response B4 we report a grid search over (τmin⁡start,τmin⁡end)(\tau_{\min}^{\mathrm{start}},\tau_{\min}^{\mathrm{end}}). The sensitivity is moderate, and replacing the rolling schedule with the best static one increases the CRPS by 2–11%.

[W3], [Q3]: SDEdit ablation.

By Proposition 2, SDEdit-style noising is the special case of the windowed inversion with ετ=∣τ˙∣1−ττ\varepsilon_\tau=\vert\dot\tau\vert\frac{1-\tau}{\tau} which tends to infinity as τ→0\tau\to0. Our parametrization uses ετ=ε∣τ˙∣(1−τ)\varepsilon_\tau=\varepsilon\vert\dot\tau\vert(1-\tau), so SDEdit corresponds to ε=1/τ\varepsilon=1/\tau. We did not run SDEdit as a separate ablation, but our tuning considered larger values of ε\varepsilon. On SQG, these consistently gave worse results, and the selected ε=0.03\varepsilon=0.03 is about two orders of magnitude below the SDEdit value, matching the result with DAISI [1] which also reported worse results for SDEdit. On SEVIR, the selected ε=2.5\varepsilon=2.5 coincides with the SDEdit value at the oldest slot (τmin⁡start=0.4\tau_{\min}^{\mathrm{start}}=0.4), indicating that SDEdit can work in this setting. At the newest slot (τmin⁡end=0.1\tau_{\min}^{\mathrm{end}}=0.1), however, SDEdit corresponds to ε=10\varepsilon=10, four times the selected value. Unlike the block smoother, which needs SDEdit to avoid inverting under the conditional prior, the windowed inversion lets τmin⁡\tau_{\min} and ε\varepsilon be chosen separately for each setting.

[1] Andrae, M., Wikingsson, E., Takao, S., Landelius, T., & Lindsten, F. (2026). DAISI: Data Assimilation with Inverse Sampling using Stochastic Interpolants. ICML 2026

[W3], [Q4]: Inconclusive SEVIR results.

We agree that the SEVIR differences are small against the reported standard deviations. These are dominated by the differences between the 10 events, which have very varying norms. To make the difference more clear, we compare them event by event in B5, which shows a consistent CRPS gain for DAWIS across all but 3 events.

[Q4]: Inference Cost Comparison.

We now report wall-clock time, network evaluations and peak memory for every method in Appendix K (B2). DAWIS is among the more expensive methods. However, the increased computational cost is well justified by the improved skill-to-compute ratio shown in Figures 34–36.

[Q4]: SEVIR Resolution.

Thank you for spotting this typo. The SEVIR domain covers 384×384384\times384 km at 3 km resolution, giving a 128×128128\times128 grid. We have corrected the text.

Dear Reviewer YwYg,

The discussion period ends in less than 24h. We believe that we have addressed all your concerns. If there are any remaining questions we'd be happy to answer them.

Regards, The authors

Reviewer_r2RSOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper extends DAISI from a single state to a window assimilation. Thanks to multitask interpolant, each state in the window can have its own flow time. The turning points τmin⁡\tau_{\min} can decide among filtering, fix-lag smoothing and block smoothing. In DAWIS-Joint, the window prior also produces the forecast. Besides, the paper also shows the update leaves πB\pi_{\mathcal B} invariant when guidance is exact for the block smoother.

Critical strengths
  1. DAWIS claims a single trained prior can handle three DA regimes: filtering, fix-lag smoothing and block smoothing. And the lagged estimate comes at very low cost from filtering.

  2. On SQG benchmark, DAWIS achieves the best CRPS and RMSE in all four obs settings against many baselines. The appendix also checks calibration, power spectrum and errors.

  3. The block smoother is supported by Prop.1, and it fixes the case where Block Gibbs can make the initialization worse in practice.

Critical weaknesses
  1. Probabilistic interpratation of the sequential update. In each cycle, a window is guided by yn−w:ny_{n-w:n}, though the input states x^n−w:n−1\hat x_{n-w:n-1} have already absorbed each and every one of these obs except the last one yny_n in earlier cycles. It is unclear to me what is the distribution the sequential update is targeting. Besides, I am curious whether guiding on the same obs for several times can introduce bias. The prop. 1 coves only the block smoother, without the filter.

  2. How the turning points are set. I didn't find in the paper about how to choose the per-slot values, given that DAWIS heavily depends on τmin⁡\tau_{\min}. The table 3 lists many undefined terms, e.g., τstartmin⁡,τendmin⁡,τinitmin⁡\tau^{\min}_{\text{start}},\tau^{\min}_{\text{end}},\tau^{\min}_{\text{init}}, and the values change between settings (τendmin⁡\tau^{\min}_{\text{end}} ranges from 0.1 to 0.4). The study on τmin⁡\tau_{\min} profile or the window size ww is also missing, which can make it hard to tell how much of the gain on Joint AR W∣1W|1 comes from windowed inversion or from per-setting tuning.

  3. Cost, and the SEVIR results. The manuscript does not report NFEs or wall-clock time. In addition, the SEVIR results vetween methods are small, next to the spread across trajectories, e.g., DAWIS filter reports 3.86±2.123.86\pm2.12 where SDA-filter is 3.79±2.253.79\pm2.25 over 10 trajectories. I also find the SQG results are pretty convincing, while the SEVIR seems less.

Questions for authors
  1. What distribution does the sequential DAWIS cycle sample from with an exact prior and exact guidance? How does it depend on τmin⁡\tau_{\min}? Can re-guiding on yn−w:n−1y_{n-w:n-1}, considering those earlier cycles have already used them, bias the filter, say, by making it overconfident? Can you perform an ablation study to see how it would change if only yny_n is used for guidance instead of the full window?

  2. How is each τmin⁡k\tau^k_{\min} obtained from the parameters in Tab. 3? How results are sensitive to τmin⁡\tau_{\min} and to ww?

  3. Could you report per-cycle NFEs and wall-clock time? Could you provide paired per-trajectory differences given that the SEVIR gains are a bit marginal?

  4. ForcingDAS also has filtering and smoothing modes, but only its fix-lag smoothing mode was compared to DAWIS. Considering this is one of the closest work, is there a reason the other two modes were discarded in comparison?

  5. The SEVIR resolution is reported as 384 in main text but 128 in appendix. Is this a typo?

  6. Appendix D.1 says the SEVIR is non-markovian, this breaks the assumption in Eq.1. Does this mean prop.1 does not cover the block smoother results for SEVIR?

3 comments

We thank the reviewer for their thoughtful questions and positive assessment of the paper. We are glad that the reviewer values our contribution of a unified DA framework and hope to address all concerns below. We refer to two common responses: A (the DAWIS target) and B (new experiments).

[W1], [Q1]: Probabilistic interpretation of the sequential update.

We appreciate the request for the DAWIS target and have added its derivation in Appendix D.1. Common response A gives the final expression and its interpretation. In short, re-guiding on already assimilated observations doesn't always incur a double count. When ε=0\varepsilon=0, the inversion removes nothing, so the re-used observations are counted more than once and the filter becomes over-confident (SSR 0.80 in the linear-Gaussian experiment). With ε>0\varepsilon>0, the inversion forgets part of the old information and the guidance restores it.

To investigate the sensitivity to the guidance set, we added an ablation in which DAWIS guides on yny_n alone (B3). For the filters, this degraded results slightly, and DAWIS remains the best method in every setting. When guiding on the full window we don't observe overconfidence as SSR > 1 accross the settings.

[W2], [Q2]: How the turning points are set.

We agree that Table 3 was not self-explanatory and thank the reviewer for pointing this out. The turning points (τmin⁡0,…,τmin⁡w)(\tau_{\min}^0,\dots,\tau_{\min}^w) are defined as a linear interpolation between τmin⁡start:=τmin⁡0\tau_{\min}^{\mathrm{start}}:=\tau_{\min}^0 and τmin⁡end:=τmin⁡w\tau_{\min}^{\mathrm{end}}:=\tau_{\min}^w which are hyperparameters. While there is more freedom than this, this restriction reduces the number of parameters. With τmin⁡start>τmin⁡end\tau_{\min}^{\mathrm{start}}>\tau_{\min}^{\mathrm{end}}, new states are inverted furthest and revised most, while older states, which have absorbed more observations, are mostly kept. The other two values only matter during the first ww cycles, while the window still reaches back to the initial condition. τmin⁡init-IC\tau_{\min}^{\mathrm{init\text{-}IC}} is the turning point of the slot holding the initial condition, and τmin⁡init\tau_{\min}^{\mathrm{init}} that of the slots before it, which start out as noise but are updated during the ww first cycles.

In common response B4 we report a grid search over (τmin⁡start,τmin⁡end)(\tau_{\min}^{\mathrm{start}},\tau_{\min}^{\mathrm{end}}). The sensitivity is moderate, and replacing the rolling schedule with the best static one increases the CRPS by 2–11. We also perform a more detailed investigation of this in the linear-Gaussian experiments (A3).

[W2]: Window size ablation.

We have performed an ablation on the window size w=2,4,6w=2,4,6 (B1). Throughout all experiments, DAWIS is the least affected by ww, and its w=2w=2-variant still outperforms all other filters using w=6w=6. We can also clearly see the degradation of Joint AR W∣1W\vert 1, which stops performing well at shorter windows.

[W3], [Q3]: Inference compute cost.

To be transparent about the inference cost of DAWIS, we have added a detailed comparison in B2. In short, the expensive DAWIS cycle is justified by its skill-to-compute ratio. We have also performed matched inference-cost and lag comparison on ForcingDAR-Pyr in B7.

[W3], [Q3]: Inconclusive SEVIR results.

We agree that the SEVIR differences are small against the reported standard deviations. These are dominated by the differences between the 10 events, which have very varying norms. To make the difference more clear, we compare them event by event in common response B5, which shows a consistent CRPS gain for DAWIS compared to all models across all but 3 events.

[Q4]: ForcingDAS filtering and smoothing modes.

Unlike DAWIS which uses a relatively short window for its prior, ForcingDAS uses the entire sequence length p(x1:n)p(x_{1:n}). This means that their full sequence smoother does not directly apply to the windowed setting we consider in this paper. However, both their filtering and fixed-lag smoothing variants can be adapted to the windowed setting, where the filter variant becomes equivalent to FlowDAS or Joint AR 1∣W1\vert W.

[Q5]: SEVIR Resolution.

Thank you for spotting this typo. The SEVIR domain covers 384×384384\times384 km at 3 km resolution, giving a 128×128128\times128 grid. We have corrected the text.

[Q6]: Non-Markovian Block Smoothing.

The point that the block-smoothing result does not directly apply to non-Markovian systems is correct, and we appreciate the chance to clarify it. The restriction to first-order Markov was made for convenience. The proof generalises directly to kk-th order Markov chains by conditioning on kk states on each side of the block, which we will clarify in the paper. Unfortunately, guarantees for truly non-Markovian systems, such as SEVIR, are difficult to obtain, and the common practice is to assume Markovianity of some finite order. In the SEVIR experiments we tried increasing the number of boundary states, but saw no increase in performance.

I thank the authors for their detailed response. They have addressed most of my concerns, and I maintain my recommendation of weak accept (3).

We appreciate that the reviewer acknowledges that most of their concerns have been addressed, and thank them for their continued support of the paper.

If there are any remaining questions or concerns, we'd be happy to answer them in the last 24h of the discussion period.

Regards, The authors

Reviewer_kQi2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DAWIS, a generative data-assimilation framework based on multitask stochastic interpolants over a short temporal window. Unlike standard filters that assimilate the newest observation while keeping earlier analyses effectively fixed, DAWIS partially inverts a window of recent states toward noise and regenerates the entire window under observation guidance. Separate turning points for each temporal slot determine how much each state is preserved or revised. The framework recovers filtering, fixed-lag smoothing, and block smoothing through different schedules, while DAWIS-Joint generates the newest state within the window and therefore avoids requiring a separate forecast model. The paper proves that its block update preserves the desired conditional posterior under exact learned conditional expectations, exact likelihood guidance, and Markov dynamics. Experiments on SQG and SEVIR report improvements in CRPS and RMSE, together with calibration, spectral, and qualitative diagnostics.

Critical strengths
  1. The method addresses a limitation of sequential generative filtering: DAWIS revises a short recent history when new observations arrive, rather than only updating the newest state. This is consequential because later observations can reveal that an earlier analysis was incorrect, and the windowed update gives the method a principled way to repair that recent history.

  2. The framework is unified and empirically thorough on the selected benchmarks: The same learned trajectory prior supports filtering, fixed-lag smoothing, and block smoothing through changes in the inference schedule. The experiments include nonlinear, sparse, noisy, and multimodal observation settings; probabilistic metrics; calibration checks; spectral diagnostics; and several relevant baselines. The consistent SQG gains make the empirical result promising.

Critical weaknesses
  1. The theoretical result does not directly apply to the SEVIR block-smoothing experiment. Proposition 1 and Corollary 1 require a Markov state-space model, exact conditional expectations, and exact partial-noise likelihood guidance. The paper explicitly states that SEVIR is non-Markovian and that the two-boundary conditioning used by the block smoother is insufficient in principle. Thus, while SEVIR performance may still be useful empirically, the paper does not establish that the SEVIR block smoother targets the claimed posterior or has the stated MCMC interpretation. This limitation is important because SEVIR is a primary experimental setting.

  2. The theory-to-practice gap and practical efficiency are insufficiently quantified: The derivations appear clear under the authors’ assumptions, but the actual method relies on approximate MMPS guidance and learned neural estimates. Important settings, including turning points and guidance strength, are tuned separately by observation scenario; the paper reports only ten evaluation trajectories; and it provides no matched inference-cost comparison. Since methods use different solver budgets and DAWIS performs window-level inversion and regeneration, the paper should report wall-clock time or neural-function evaluations, memory, and accuracy-versus-compute curves. Without these, it is more difficult to assess robustness and practical scalability.

Questions for authors
  1. For the non-Markovian SEVIR setting, what distribution does the two-boundary block smoother approximate? Please evaluate longer conditioning contexts or provide evidence that the learned seven-frame prior makes conditioning on two boundaries adequate.

  2. Can the authors quantify the theory-to-practice gap from approximate MMPS guidance? In particular, how sensitive are the results to the Gaussian approximation, the neglected derivative of the conditional covariance, and the numerical handling of the estimated covariance?

  3. Please provide the complete hyperparameter-selection procedure: validation trajectories, search ranges, selection criterion, and the tuning budget for each baseline. Can the authors also report variance across independently trained model seeds?

  4. Please provide matched inference-cost comparisons: wall-clock time or neural-network evaluations per assimilated state, peak memory, and accuracy-versus-compute curves. Does DAWIS retain its advantage under matched budgets?

  5. The vector-valued flow-time schedule τ\tau, and thus the vector coefficients ατ\alpha_\tau and βτ\beta_\tau, are central to DAWIS. Please clarify in the main text which components share a flow time, how per-slot flow times are sampled during training, and how turning points and forward/reverse schedules are selected at inference. Are these theoretically prescribed or tuned hyperparameters? Please also report sensitivity to these choices.

2 comments

We thank the reviewer for their time and thorough review. We are especially grateful that the reviewer highlights our contribution towards solving the fixed-history limitation of existing generative filters. Below we address the questions and concerns. We refer to two common responses: A (the DAWIS target) and B (new experiments, items B1–B5).

[W1], [Q1]: Non-Markovian Block Smoothing.

The point that the block-smoothing result does not directly apply to non-Markovian systems is correct, and we appreciate the chance to clarify it. The restriction to first-order Markov was made for convenience. The proof generalises directly to kk-th order Markov chains by conditioning on kk states on each side of the block, which we will clarify in the paper. Unfortunately, guarantees for truly non-Markovian systems, such as SEVIR, are difficult to obtain, and the common practice is to assume Markovianity of some finite order. In the SEVIR experiments we tried increasing the number of boundary states, but saw no increase in performance.

[W2], [Q2]: Biased guidance methods.

While we considered MMPS in this paper, this is simply one of the possible choices. Future developments of better guidance methods are directly applicable to DAWIS. With MMPS, the default strength ζ=1\zeta=1 was sufficient for Noisy, Sparse and Multimodal while SEVIR used ζ=2.5\zeta=2.5. For Saturating we used ζ=20\zeta=20 for the filter and ζ=150\zeta=150 for the block smoothers. With arctan observations and σobs=0.01\sigma_{\mathrm{obs}}=0.01, the observation Jacobian is nearly zero in the saturated region, so the linearised MMPS likelihood gradient is far too weak and the strength compensates for this. For the baselines, the guidance strength was tuned per experiment, with performance at times being sensitive to this value.

Our theoretical results for DAWIS assume exact guidance, so in the nonlinear experiments the guidance error and the bias of the sequential scheme cannot be separated. To separate them, we added an experiment on a linear-Gaussian system, where the generative model is available in closed form and MMPS is exact (A3).

[Q3]: Hyperparameter tuning.

For DAISI, FlowDAS, LETKF and EnSF, we used the values from [1]. For the rest of the baselines (including DAWIS), the tuning was done on a single validation trajectory using an active exploration strategy rather than exhaustive grid searches. The tuning budget depended on the number of hyperparameters and sensitivity of the method, but significant effort was spent trying to make the baselines work well. The tuning was primarily done on the sparse setting, with additional tuning for each experiment. The search ranges were wide, with increasing resolution as the range got smaller. The search was stopped when the difference between consecutive values was inconclusive. Due to our approach, we don't have an exact tuning budget for each baseline.

[1] Andrae, M., Wikingsson, E., Takao, S., Landelius, T., & Lindsten, F. (2026). DAISI: Data Assimilation with Inverse Sampling using Stochastic Interpolants. ICML 2026

[Q3]: Model seeds.

We have repeated the model training using three additional random seeds and included the results in Appendix I. These results indicate that the choice of training seed for the backbone model has no significant impact on the evaluation results.

[W2], [Q4]: Matched inference-cost comparisons.

To be transparent about the inference cost of DAWIS, we have added a detailed comparison in B2. In short, the expensive DAWIS cycle is justified by its skill-to-compute ratio. We have also performed matched inference-cost and lag comparison on ForcingDAR-Pyr in B7.

[Q5]: Choice of ατ\alpha_\tau, βτ\beta_\tau, τt\tau_t schedules.

DAWIS uses the multitask interpolant prior with a single flow time τk\tau_k for each state xkx_k in our window x0,…,xwx_0,\dots,x_w. This allows individual control of the noise level in each state, which we leverage for windowed inversion and conditioning on frozen states. During training, the interpolant uses the standard linear schedule βτ=τ\beta_\tau=\tau, ατ=1−βτ\alpha_\tau=1-\beta_\tau, sampling a different τk∼U([0,1])\tau_k\sim U([0,1]) for each state xkx_k. For regular flow matching, the training schedule also dictates the inference schedule. However, with the multitask interpolant, this is instead controlled by the schedule τt\tau_t which does not necessarily have to be a linear function of tt. We will clarify this further in the main paper.

The turning points (τmin⁡0,…,τmin⁡w)(\tau_{\min}^0,\dots,\tau_{\min}^w) are defined as a linear interpolation between τmin⁡start:=τmin⁡0\tau_{\min}^{\mathrm{start}}:=\tau_{\min}^0 and τmin⁡end:=τmin⁡w\tau_{\min}^{\mathrm{end}}:=\tau_{\min}^w which are hyperparameters. While there is more freedom than this, this restriction reduce the number of tunable parameters. We report a grid search over these parameters in B4. We also perform a more detailed investigation of this in the linear-Gaussian experiments (A3).

Dear Reviewer kQi2,

The discussion period ends in less than 24h. We believe that we have addressed all your concerns. If there are any remaining questions we'd be happy to answer them.

Regards, The authors

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

This paper introduces DAWIS, a generative data assimilation framework built on a multitask stochastic interpolant over short state windows. Each assimilation cycle partially inverts previous estimates and regenerates the window under observation guidance, allowing new observations to revise past states. Different initialization and turning-point choices support filtering, fixed-lag smoothing, and forecasting without an external model. An overlapping-block refinement extends the same prior to full-trajectory smoothing. Experiments on SQG and SEVIR report improvements over several filtering and smoothing baselines.

Critical strengths

Windowed inversion revises historical estimates, including the oldest conditioning state, without regenerating everything from scratch. This directly addresses the fixed-history limitation and provides filtering and lagged-smoothing estimates within the same cycle.

Critical weaknesses
  1. The online update’s posterior interpretation is unclear. Previously assimilated samples undergo unconditional inversion and are guided again by overlapping observations. Unlike the block smoother, this update lacks a justification that it targets the intended posterior. Possible bias from retaining and reusing past evidence is therefore not distinguished from guidance approximation.
  2. This paper reported 100 steps and six observations.for each screen. Without matched budgets and look-ahead, or an ablation isolating revision of the oldest state, the benefit of windowed inversion remains insufficiently established.
Questions for authors
  1. Clarification of the target distribution. With an exact prior and guidance, what distribution does the online update produce, and how does it avoid incorrectly reweighting previously assimilated evidence?

  2. Evidence for the proposed mechanism. Does DAWIS-Joint retain its advantage with matched inference cost and look-ahead? How much does performance change when only the oldest state is frozen?

3 comments

We thank the reviewer for their time and positive feedback. We are glad that the reviewer recognizes the benefit of windowed inversion for revising historical estimates. Below we address the questions and concerns. We refer to two common responses: A (the DAWIS target) and B (New experiments).

[W1], [Q1]: Clarification of the target distribution.

We agree on the importance of this and have added a derivation of the DAWIS target in Appendix D.1. Please see common response A for the final expression and its interpretation. In short, re-guiding on already assimilated observations doesn't always incur a double count. When ε=0\varepsilon=0, the inversion removes nothing, so the re-used observations are counted more than once and the filter becomes over-confident (SSR 0.80 in the new linear-Gaussian experiment (A3). With ε>0\varepsilon>0, the inversion forgets part of the old information and the guidance restores it.

To investigate the sensitivity to the guidance set, we added an ablation in which DAWIS guides on yny_n alone (B3). This had a moderate effect on the filter performance, but degraded results for the smoother in settings with a larger ε\varepsilon.

[W2], [Q2]: Inference-cost comparisons.

To be transparent about the inference cost of DAWIS, we have added a detailed comparison in Appendix K (B2). DAWIS is among the more expensive methods, since each cycle inverts and then regenerates the whole window. However, the increased computational cost is justified by the improved skill-to-compute ratio shown in Figures 34–36.

We have experimented with increasing the number of steps for the baseline methods as part of the tuning process. The reported results for the baselines are for the best setting on the tuning trajectory where more steps have led to either the same or worse performance. For example, rerunning ForcingDAS-Pyr with 200 steps per cycle, which costs more than a DAWIS cycle, does not significantly improve the result for any setting (B7).

[W2], [Q2]: Window size ablation.

For look-aheads, we chose to use a same size prior across baselines, implying that methods with a frozen anchor (Joint AR, ForcingDAS-Pyr) saw fewer observations each cycle. To verify that the difference in lag does not affect the performance, we have performed an ablation on the window size w=2,4,6w=2,4,6 (B1). Throughout all experiments, DAWIS is the least affected by ww, and its w=2w=2-variant still outperforms all other filters using w=6w=6.

[W2], [Q2]: Benefit of revising the oldest state.

The benefit of revising the oldest state can be seen from the gap in performance between DAWIS and Joint AR W∣1W\vert 1, which uses a frozen anchor. Further, to show that this holds more generally, we also perform a linear-Gaussian experiment, comparing DAWIS to the true posterior. There, we see that freezing the oldest state consistently yields a lower performance, highlighting the benefit of windowed inversion. For more details on this, we refer to common response A3.

Dear Reviewer FCVH,

The discussion period ends in less than 24h. We believe that we have addressed all your concerns. If there are any remaining questions we'd be happy to answer them.

Regards, The authors

Thanks for the detailed rebuttal. Given the added experiment results and other reviews, I believe the current ratings reflect my evaluation of this paper.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Common response A: What does DAWIS sample from?

We thank the reviewers for their feedback regarding the ambiguity of what DAWIS samples from and appreciate the chance to clarify this. To address this concern, we have carefully derived the target for DAWIS, which has been added to Appendix D.1. Further down we present the final expression and interpret its parts. Before that, we give some thoughts on unbiased DA in general.

A1: No practical method for high-dimensional nonlinear DA samples from the true target

We want to be transparent that DAWIS is biased in principle, and to position this within the field. To the best of our knowledge, particle filters and related interacting-particle methods are the only approaches that asymptotically sample from the true filtering or smoothing distribution in nonlinear, non-Gaussian settings, and they suffer from the curse of dimensionality in the regimes we consider. All practical data assimilation methods for high-dimensional nonlinear problems, generative or EnKF-based, trade asymptotic consistency for bias in order to remain tractable, and this applies to every baseline in the paper.

Specifically, it applies to all DA methods based on guidance of a generative model, which is the category to which DAWIS belongs. All such methods have two types of bias: (1) The bias coming from non-exact guidance. (2) The bias coming from temporal dependencies between the current observation and past states.

The first type can in principle be mitigated by advances in guidance and inference-time control in general. We have used MMPS as a generic choice of guidance method since we have found it to provide a good tradeoff between simplicity and performance, but the DAWIS framework is by no means limited to using MMPS. As new and better guidance methods are developed they can be integrated in the framework.

The second type of bias is what stands out in the context of dynamical systems, and different generative DA methods handle this in different ways. The key merit of DAWIS is to combine the inversion strategy of DAISI with windowed sampling, to reap the benefits of both—a combination that is made possible by the nontrivial application of multitask SI applied to windows sampled from the stationary distribution. We contrast this with alternative generative DA methods below.

FlowDAS and Joint AR 1∣W1\vert W target ∏s≤np(xs∣ys,xs−1)\prod_{s\le n}p(x_s\mid y_s,x_{s-1}). From a particle-filtering perspective, this is the product of locally optimal proposals without the resampling weights p(ys∣xs−1)p(y_s\mid x_{s-1}). Its marginals can differ substantially from p(xn∣y1:n)p(x_n\mid y_{1:n}), and for strong dependence between state variables it can collapse to the prior.

Joint AR W∣1W\vert 1 moves the anchor backward in time, giving the interior more freedom to move towards the observations. However, it still freezes the anchor xn−wx_{n-w} and drops the weight p(yn∣xn−w,yn−w+1:n−1)p(y_n\mid x_{n-w},y_{n-w+1:n-1}), risking the same problems as Joint AR 1∣W1\vert W. This issue is particularly severe for short windows, which in practical (high-dimensional) cases might be the only option due to the high memory requirements of regenerating long windows.

SDA-filter regenerates the anchor together with the interior, but replaces the previous analysis by a pseudo-observation of tuned strength without any guarantees.

ForcingDAS-Pyr has no clear target distribution. Each frame enters the pyramid at pure noise and is guided by its own observation while it descends. Later observations also act on earlier frames at every step, but with fading effect as the earlier frame approaches a clean sample. Once a frame is finalised, it enters the fixed conditioning and acts as a frozen anchor, inheriting the bias of Joint AR W∣1W\vert 1.

SDA targets the joint smoothing distribution p(x1:n∣y1:n)p(x_{1:n}\mid y_{1:n}). However, the pseudo-blanket approximation that makes it scalable introduces a bias that remains even under exact guidance. This can manifest in form of difficulties in multimodal settings.

EnKF is the only baseline with any form of guarantees, but those hold only in the near-Gaussian regime and degrade in nonlinear, non-Gaussian settings. Its practical variants, such as LETKF, are motivated by empirical performance rather than stronger guarantees.

In practical settings without guarantees, methods have to be judged by thorough experimental validation. Our results demonstrate that DAWIS works well under a range of complex observation operators, even with imperfect guidance methods. Still, we believe transparency is important and provide the DAWIS target below.

AuthorsOpenReview ↗

A2: What does DAWIS sample from?

A2: What does DAWIS sample from?

For easier presentation, we restrict the derivation to ε=0\varepsilon=0. Assume that the forecast is absolutely continuous with respect to the windowed prior. We can then relate the DAWIS cycle to the fixed-lag posterior p(xn−w:n∣y1:n)p(x_{n-w:n}\mid y_{1:n}) under exact prior and guidance by viewing DAWIS as a proposal and deriving its importance weights. Given an input x^n−w:n∼p(xn−w:n∣y1:n−1)\hat x_{n-w:n}\sim p(x_{n-w:n}\mid y_{1:n-1}), let xn−w:nx_{n-w:n} denote the output from the DAWIS cycle with inverted latent zτmin⁡z_{\tau_{\min}}. Then, the weights for every choice of τmin⁡\tau_{\min} are

wn=fn−w(xn−w)fn−w(x^n−w)⏟(i) anchor displacement⋅p(yn−w:n∣zτmin⁡)⏞(ii) incomplete guidance∏k<np(yk∣x^k)⏟(iii) double countingw_n=\underset{\text{(i) anchor displacement}}{\underbrace{\frac{f_{n-w}(x_{n-w})}{f_{n-w}(\hat x_{n-w})}}}\cdot\frac{\overset{\text{(ii) incomplete guidance}}{\overbrace{p(y_{n-w:n}\mid z_{\tau_{\min}})}}}{\underset{\text{(iii) double counting}}{\underbrace{\prod_{k<n}p(y_k\mid\hat x_k)}}}

where p(y∣zτmin⁡)=E[p(y∣z1)∣zτmin⁡]p(y\mid z_{\tau_{\min}})=\mathbb E[p(y\mid z_1)\mid z_{\tau_{\min}}], and

fn−w(xn−w)=p(xn−w∣y1:n−w−1)P∞(xn−w)f_{n-w}(x_{n-w})=\frac{p(x_{n-w}\mid y_{1:n-w-1})}{\mathbb P_\infty(x_{n-w})}

is the density ratio of the anchor's predictive distribution with respect to the prior. The weight is exactly one when DAWIS samples from p(xn−w:n∣y1:n)p(x_{n-w:n}\mid y_{1:n}) and its factors are the sources of bias. The derivation can be found in Appendix D.1 which also treats the case of only guiding toward the latest observation.

Factor (i): revising the anchor vs freezing it.

The forecast information that survives the cycle is the anchor's density ratio, evaluated at the input x^n−w\hat x_{n-w} instead of the output xn−wx_{n-w}. Freezing the anchor, τmin⁡0=1\tau_{\min}^0=1, makes this factor one, but then the information that yny_n carries about xn−wx_{n-w} cannot enter, as in Joint AR W∣1W\vert 1. A partial inversion of the anchor is what lets DAWIS revise it, at the price of a displacement that grows with the inversion depth.

Factor (ii): not starting the guidance in τmin⁡=0\tau_{\min}=0.

The numerator is the likelihood of the window observations seen from the latent. At τmin⁡=1\tau_{\min}=1 it is the full likelihood ∏kp(yk∣xk)\prod_k p(y_k\mid x_k), nothing has been assimilated and wnw_n reduces to the bootstrap-filter weight p(yn∣xn)p(y_n\mid x_n). At τmin⁡=0\tau_{\min}=0 it is constant, the guided flow has injected all the observation information. In between it measures the observation information not injected on [τmin⁡,1][\tau_{\min},1], so a larger τmin⁡\tau_{\min} improves factor (i) and worsens factor (ii).

Factor (iii): re-insertion of previously assimilated observations.

The denominator reflects the re-insertion of previously assimilated observations. At ε=0\varepsilon=0, none of the information is removed by the inversion, and the past observations are counted twice, once in the input and once through the guidance. How much the second count matters furthermore depends on τmin⁡\tau_{\min}. Intuitively, for τmin⁡\tau_{\min} close to 1 there is "less time" to incorporate the observations through guidance. However, with ε>0\varepsilon>0, the inversion forgets some of the forecast information, weakening the dependence on past observations, which is what requires us to re-insert them during assimilation. At the end of the day, the interplay between τmin⁡\tau_{\min} and ε\varepsilon becomes a trade-off between guidance strength ("double counting" can be seen as cranking up guidance strength) and dependence on the anchor state, which carries information about pre-window observations to the current state. We believe that all guidance-based DA methods have this trade-off. While some methods rely heavily on an explicit guidance strength that is manually tuned, DAWIS shifts this design choice to the choice of τmin⁡\tau_{\min} and ε\varepsilon.

Although the bias has a more complicated form than that of Joint AR W∣1W\vert 1, neither method is unbiased, and what matters is how well it works in practice. We hope that this derivation addresses the reviewers' concerns and are happy to answer any further clarifications.

AuthorsOpenReview ↗

A3: Linear-Gaussian comparison.

A3: Linear-Gaussian comparison.

To isolate the bias of the sequential scheme, we run DAWIS on a linear-Gaussian state-space model and compare it with the exact fixed-lag (Kalman/RTS) posterior. We consider damped rotations in d=16d=16 with every second component observed and R=0.3IR=0.3I, a window of 4 states, and a high and low model noise level qq. The prior is Gaussian, so the velocity field is known in closed form and MMPS guidance is exact. The only remaining error is therefore the bias of the sequential scheme itself. Every stage of a cycle is affine-Gaussian, so we propagate the ensemble mean and covariance exactly instead of sampling. The turning points are linear from the anchor xn−3x_{n-3}, at τmin⁡start\tau_{\min}^{\mathrm{start}}, to the newest state xnx_n, at τmin⁡end\tau_{\min}^{\mathrm{end}}. Figure 42 in Appendix P shows the steady-state KL to the exact window posterior over a (τmin⁡start,τmin⁡end,ε)(\tau_{\min}^{\mathrm{start}},\tau_{\min}^{\mathrm{end}},\varepsilon) grid, for guidance on all window observations and on the last one, compared with Joint AR W∣1W\vert 1. The table summarises the best settings.

MethodKL (q=0.05q=0.05)SSRRMSE/RTSKL (q=0.01q=0.01)SSRRMSE/RTS
Kalman/RTS (exact)011011
DAWIS, guide on yny_n (best)0.581.011.0350.110.981.011
DAWIS, guide on yn−3:ny_{n-3:n} (best)1.361.071.0720.750.921.031
DAWIS, guide on yn−3:ny_{n-3:n}, ε=0\varepsilon=06.360.801.1561.570.801.046
Joint AR W∣1W\vert 11.561.011.0942.251.011.129

1. Re-guiding at ε=0\varepsilon=0 double counts.

A deterministic inversion removes nothing from the retained states, so the re-used observations are counted twice. This is factor (iii) of the DAWIS bias, and it shows up as an SSR of 0.80 in both regimes. With ε>0\varepsilon>0, the inversion forgets part of the old information and the guidance restores it.

2. Retaining old information works better than removing and re-inserting when guidance is exact.

Guiding on yny_n only is the best variant in both regimes. It keeps the old information in the latent instead of removing and re-inserting it. Nothing restores what ε\varepsilon erases, so it needs a much smaller ε\varepsilon (0.01-0.03 vs 0.1-1). A small positive ε\varepsilon is still needed, because it supplies the spread that a deterministic cycle lacks. With approximate guidance, the advantage might not be as clear as we see in the SQG experiments.

3. Revising the anchor is what improves on Joint AR W∣1W\vert 1.

Joint AR W∣1W\vert 1 regenerates the interior from noise, so old information is fully removed before it is re-inserted. Its spread is calibrated, and its whole error lies in the frozen anchor. Every best DAWIS schedule has τmin⁡start\tau_{\min}^{\mathrm{start}} between 0.6 and 0.9, so the anchor is always partially revised. The gain is largest at low model noise (KL 0.11 vs 2.25), where the uncertainty carried over from the previous analysis dominates, and a frozen anchor is most restrictive. There, the region of schedules that beat Joint AR is also broad.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.