acceptodds
Under review as a conference paper at ICLR 2027

AECSF: Adaptive Ensemble Conditional Score Filtering for High-Dimensional Nonlinear Data Assimilation

Abstract

Bayesian state estimation for high-dimensional nonlinear dynamical systems entails a fundamental tension between statistical fidelity and computational tractability, as particle weights can collapse, while Gaussian ensemble updates can miss non-Gaussian posterior structure. Score-based diffusion filters offer a sampling-based alternative, but existing training-free score filters often rely on heuristic likelihood corrections, which can compromise posterior accuracy by neglecting uncertainty about the system state associated with each noisy reverse particle. To address these issues, we propose AECSF, a training-free adaptive ensemble conditional score filter. AECSF constructs an analytically tractable score estimator from the conditional Tweedie identity, which recasts noisy posterior score estimation as estimating the conditional mean of the system state given a noisy reverse particle and the observation. To estimate these conditional means efficiently, AECSF employs a shared adaptive weighted proposal ensemble, while particle-specific conditional weights yield an estimate for each noisy reverse particle without separate proposal sampling. The proposal ensemble is updated using reverse-particle information within the same reverse-diffusion run to improve conditional-mean estimation. Theoretically, we characterize when a fixed weighted proposal measure yields the exact noisy posterior score. Under stated assumptions, we establish a bound relating conditional-mean estimation errors to reverse-sampling endpoint error. Numerical experiments demonstrate that AECSF improves the accuracy of posterior sampling and nonlinear filtering in high-dimensional problems with limited forecast ensembles.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Data assimilation, Nonlinear filtering, Score-based diffusion

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_cMt8OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses the filtering problem based on ensemble score filtering framework. For score-based filtering methods, evaluation of conditional score for intermediate reverse samples is necessary for guiding the sample generation in the update step. The authors have proposed a method to evaluate the conditional mean E[Z∣Xτ=x,Yn=yn]\mathbb{E}[Z|X_{\tau}=x, Y_n = y_n], which is necessary to obtain the score function using conditional Tweedie identity. To compute the conditional mean in a robust way, they propose to use sample approximation: conditional mean is computed as the weighted average of the proposal ensemble particles. The proposal particles are moved to well-sample the region of interest, i.e., near reverse particles. Experiments show assimilation performance comparable to other training-free filtering approaches such as LETKF and IEnSF.

Critical strengths
  • The use of ensembles to estimate the score function seems sound.
Critical weaknesses
  • The main framework of the score-based filtering is already established in the ensemble score filter. So the contribution is in the new conditional score estimation method based on an ensemble. The idea of taking samples to robustly compute some quantity is fairly straightforward and seems incremental to me.
  • The use of multiple samples for evaluating score is precedented by Binder et al. 2026. Although it is concisely discussed at the final four lines of Section 2, the comparison against this method and Binder's method is not very clear (for example, can you explain your method with the KDE language, or vice versa?). Considering the similarity, I would like to know in detail what made the significant difference in the performance of Lorenz96 results (Table 1).
  • Since the experiments do not include machine learning component, it is hard to evaluate the contribution as a machine learning method. A new guidance method could be of significant importance if strong, but the experimental results target training-free filtering problems. Empirical evidence that highlights significance for machine learning community is lacking.
Questions for authors
  • What made the significant performance difference between AECSF and Binder et al. in the Lorenz96 experiment?
  • Can AECSF guide the sample generation for a diffusion model-based data assimilation? What would the performance be?
6 comments
AuthorsOpenReview ↗

Response to Reviewer cMt8 (1/4)

We sincerely thank you for carefully reviewing our manuscript and for your specific feedback on AECSF's novelty and relevance to machine learning. Your questions concern AECSF's substantive contribution relative to existing score-based filters and KDE methods, and the research value of training-free conditional inference. We first explain the significance of this research direction, then address these questions through methodological comparisons, additional controlled experiments, and the learned-prior extension.

W3. Research significance of score-based filtering and generative data assimilation

Our work lies at the intersection of Bayesian inference and generative modeling, addressing sequential state estimation and uncertainty quantification for high-dimensional physical systems. The ICLR 2027 call for papers includes Bayesian methods, sampling, uncertainty quantification, generative models, and physical-science applications. For these problems, advances in posterior inference and sampling can constitute methodological contributions without introducing a new learning module.

In data assimilation, incomplete, noisy observations must be combined with forecasts to infer states and their uncertainty. Latent-EnSF (Si and Chen, ICLR 2025) combines learned state and observation representations with ensemble score filtering; LD-EnSF (Xiao et al., ICLR 2026) learns latent dynamics for sparse observations. FlowDAS (Chen et al., NeurIPS 2025) learns stochastic transitions and incorporates observations step by step. Score-based Data Assimilation (Rozet and Louppe, NeurIPS 2023) learns a trajectory prior and introduces the observation model only at inference.

These approaches share a conditional-inference question: how should a generative prior be updated to produce reliable samples consistent with current observations? Diffusion Posterior Sampling (Chung et al., ICLR 2023) studies approximate posterior sampling for noisy inverse problems. Twisted Diffusion Sampler (Wu et al., NeurIPS 2023) uses SMC and weighted particles for conditional diffusion sampling. These works highlight the value of improving conditional-sampling algorithms even when the generative model is already trained.

AECSF addresses this conditional-inference problem by adapting a shared proposal ensemble to estimate the conditional scores required for reverse diffusion. Its score estimator is training-free, and the learned-prior experiment in Q2 demonstrates how this inference procedure can be combined with a pretrained generative model.

AuthorsOpenReview ↗

Response to Reviewer cMt8 (2/4)

W1. AECSF: a conditional-score estimation framework with shared-proposal adaptation

We propose AECSF, a conditional-score estimation framework that couples shared-proposal adaptation with reverse sampling. As reverse particles move during denoising, the shared ensemble must support score estimation at their changing locations.

EnSF (Bao et al., 2024, Section 3.2) acknowledges that its damped likelihood-gradient guidance may not yield the exact posterior score, leaving efficient correction open. EnFF (Transue et al., 2026, Section 4.2) analyzes Monte Carlo guidance with fixed forecast particles. Its output distribution approaches the likelihood-reweighted empirical posterior of a bootstrap particle filter as terminal noise vanishes.

AECSF uses the observation likelihood directly in conditional-mean estimation. Tempered weights gradually incorporate the likelihood; Langevin moves use forecast-prior and likelihood gradients to relocate proposals. Guided ancestor selection combines target weights with each proposal’s compatibility with current reverse particles under the Gaussian noising kernel. AECSF aggregates these contributions across reverse particles and applies importance correction for the changed selection probabilities. The updated ensemble provides particle-specific conditional means, which the conditional Tweedie identity converts into scores for the next reverse updates.

Within AECSF, we held the reverse sampler and proposal-movement schedule fixed, comparing standard SMC resampling with reverse-particle-guided resampling and its importance correction. L96 RMSE decreased from 0.1708 to 0.1531 and CRPS from 0.0840 to 0.0776, with both metrics improving in all ten paired runs. This comparison quantifies the additional filtering benefit of guided proposal allocation.

AuthorsOpenReview ↗

Response to Reviewer cMt8 (3/4)

W2. AECSF and Binder in KDE language

The distinction can be seen directly in the posterior representations and score calculations. Below, ϕΣ\phi_\Sigma denotes a zero-mean Gaussian density with covariance Σ\Sigma. Binder et al. (2026, Section 3.2.3) first construct a joint KDE from paired forecast-state and observation samples (zm,y~m)(z_m,\widetilde y_m), then condition on the observation yy. Let HH and GG be the state- and observation-kernel covariance matrices, and σB,i2\sigma_{B,i}^2 the additive noising variance at the current stage. Their posterior approximation and score after noising are

p_B(z\mid y) &\propto\sum_m\phi_H(z-z_m)\phi_G(y-\widetilde y_m),\\\\ s_{B,i}(x;y) &=\nabla_x\log\\!\left[ \sum_m\phi_G(y-\widetilde y_m) \phi_{H+\sigma_{B,i}^2I}(x-z_m)\right]. \end{aligned}$$ The observation thus changes mixture weights through the observation kernel, and the score follows by analytically differentiating the noised Gaussian mixture's log density. Forecast centers remain fixed within one analysis, but each component's contribution to the score still varies with the evaluation point $x$. AECSF instead uses the original observation likelihood $L$ directly. At stage $i$, shared proposals $z_i^m$ and integration weights $W_{i,m}$ approximate the reference distribution at tempering level $\rho_i$. Incorporating the residual likelihood $L^{1-\rho_i}$ gives an approximate posterior measure for the current score-estimation step: $$\begin{aligned} \widehat p_i(\mathrm dz\mid y) &\propto\sum_m W_{i,m}L(y\mid z_i^m)^{1-\rho_i} \delta_{z_i^m}(\mathrm dz),\\\\ \widehat s_i(x;y) &=\frac{a_i\sum_m c_{i,m}(x,y)z_i^m-x}{\sigma_i^2},\\\\ c_{i,m}(x,y) &\propto W_{i,m}L(y\mid z_i^m)^{1-\rho_i} \phi_{\sigma_i^2I}(x-a_i z_i^m). \end{aligned}$$ Here, $\delta_z$ denotes a point mass at $z$, and $c_{i,m}$ is normalized over the proposals. The coefficients $a_i=a_{\tau_i}$ and $\sigma_i=\sigma_{\tau_i}$ specify AECSF's noising at the current stage. The weighted average $\sum_m c_{i,m}z_i^m$ estimates the conditional mean given both the observation $y$ and reverse particle $x$. The score expression follows from the conditional Tweedie identity and gives the unclipped score with current proposal locations and integration weights held fixed. Appendix D.4 presents this KDE comparison: Binder conditions a joint KDE with fixed forecast centers through its observation kernel. AECSF uses the original likelihood and continually adapts shared proposals during denoising through guided selection, Langevin moves, and weight updates. The shared proposals supply integration samples for score estimation, while final analysis samples are produced by reverse sampling. ### Q1. What explains the large L96 performance gap? To test how much the conditioning difference in W2 explains the L96 gap, we added **Enhanced Binder** alongside original Binder and AECSF. It retains a forecast-state KDE but replaces observation-kernel conditioning with numerical conditioning using the original likelihood: $$p_{\mathrm{enh},h}(z\mid y)\propto L(y\mid z)\frac1M\sum_m\phi_{H_h}(z-z_m).$$ Here, $H_h$ is the forecast KDE kernel covariance. Under the original arctangent likelihood, numerical conditioning updates both mixture weights and within-kernel distributions before direct sampling. We use 20 members, observation-noise SD 0.05, and the same preselected bandwidth $h=0.5$ at both dimensions. Entries are means of three paired runs; SSR is the spread–skill ratio (ideal value 1). | Dimension | Method | RMSE | CRPS | SSR | |---|---|---:|---:|---:| | 10 | Original Binder | 7.801 | 6.057 | 2.303 | | 10 | Enhanced Binder | 0.495 | 0.228 | 1.244 | | 10 | AECSF | **0.140** | **0.083** | 1.178 | | 10,000 | Original Binder | 5.128 | 4.137 | 0.000193 | | 10,000 | Enhanced Binder | 1.431 | 0.759 | 0.042 | | 10,000 | AECSF | **0.154** | **0.078** | 0.904 | At $d=10{,}000$, Enhanced Binder **closes 74.3% of the original Binder–AECSF RMSE gap**, supporting the importance of observation conditioning. It remains severely underdispersed: median component-weight ESS falls from **7.3–7.7** at $d=10$ to **approximately one** at $d=10{,}000$. Multiplying coordinatewise evidence can concentrate global weights on one component, reducing between-component variability in recursive filtering. AECSF uses a regularized prior fitted to the entire ensemble to support new proposal locations. The methods differ in forecast representation, conditioning, and proposal evolution. W1's internal control measures the added benefit of guided selection and its correction. Table 1 includes both Binder variants; Section 5.2 explains the comparison; Appendix D.4 gives paired results, the method definition, and weight diagnostics.
Reviewer_g7e8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Adaptive Ensemble Conditional Score Filtering (AECSF), a training-free method for nonlinear data assimilation. The method uses the conditional Tweedie identity to estimate posterior scores from a weighted proposal ensemble. It further updates the proposals through reverse-particle-guided resampling and Langevin movement. Experiments on Gaussian problems, Lorenz-96, and Kuramoto–Sivashinsky demonstrate improved filtering accuracy over several score-based baselines, with promising computational efficiency.

Critical strengths
  1. Well-motivated approach. Using the conditional Tweedie identity to estimate posterior scores is a reasonable and principled approach. The adaptive proposal ensemble provides a way to improve score estimation without training a neural network.
  2. Computational efficiency. The shared proposal ensemble allows efficient parallel computation across reverse particles. The method is well suited for GPU implementation and shows promising analysis-time efficiency, particularly as the ensemble size increases.
Critical weaknesses
  1. Limited clarity of the method. The overall method is difficult to follow, particularly the interaction between the proposal ensemble and reverse particles. A schematic figure illustrating the full algorithm, including proposal updates and score estimation, would make the paper much easier to understand.
  2. Unclear practical advantage over LETKF. LETKF achieves lower RMSE and CRPS than AECSF on Lorenz-96. Although AECSF is reported to be approximately 31 times faster, LETKF is naturally well suited for parallel computation. Given the small ensemble size and localization radius, the reported LETKF analysis time of 9.55 seconds seems relatively high. It is unclear whether the LETKF implementation is sufficiently optimized, making the practical efficiency advantage of AECSF difficult to assess.
Questions for authors
  1. Have the authors directly evaluated the accuracy of the adaptive proposal ensemble? Does reverse-particle-guided resampling improve the proposal distribution approximation or reduce conditional-score estimation errors compared with standard likelihood-based resampling?
  2. Could the authors evaluate AECSF under different observation densities? Since the Langevin proposal updates rely on a diagonal Gaussian approximation to the forecast distribution, how does the method perform when observations are sparse and spatial correlations become more important?
  3. Have the authors evaluated forecast accuracy after cycling data assimilation? Could the authors compare forecast RMSE at different lead times to assess whether improvements in analysis accuracy also translate into better forecasts?
  4. Could the authors clarify how the LETKF implementation is parallelized on the GPU? Are the local matrix operations performed in batches, and how much of the reported speedup comes from algorithmic efficiency rather than implementation differences?
5 comments
AuthorsOpenReview ↗

Response to Reviewer g7e8 (1/3): Method schematic and LETKF comparison

Thank you for the specific suggestions on clarity, computational efficiency, and empirical validation. We have added explanations and experiments addressing how the two ensembles interact and the practical effects of this design, and have remeasured the parallel LETKF runtime. Below we address your questions and describe the corresponding manuscript revisions.

W1. Interaction between the proposal ensemble and reverse particles

The proposals represent possible clean system states. The reverse particles are noisy states at which conditional means and scores are evaluated, and their final states form the analysis ensemble. All reverse particles share proposal locations and integration weights. However, a proposal contributes differently to the estimates needed by different reverse particles, so each reverse particle computes its own conditional coefficients. Figure 1 shows the interaction in three stages. First, the current reverse particles guide shared ancestor selection. Next, the selected proposals receive weight corrections and are moved while the reverse particles remain fixed. Each reverse particle then uses the updated proposals to compute its own conditional mean and score and take a reverse step. The revised method overview follows this sequence, with formulas placed beside the relevant explanation.

W2 / Q4. LETKF accuracy, GPU parallelization, and the reported speedup

The original LETKF implementation performed its computations on the GPU but dispatched independent local analyses sequentially from Python. To address your concern about parallel efficiency, we scheduled local analyses across eight CUDA streams and reused cached workspaces and CUDA graphs to reduce execution overhead. Remeasurement under the original experimental settings reduced complete analysis time from 9.5530 seconds to 0.5093±0.00140.5093\pm0.0014 seconds per cycle. To verify that parallel optimization preserved the original numerical results, we reran all ten original L96 seeds for 200 cycles each using the archived truth, observations, initial ensembles, and numerical settings. The parallel code reproduced the scalar implementation's outputs and all 30 archived checkpoint ensembles elementwise, with the original RMSE and CRPS recovered to rounding precision.

After this parallel optimization, AECSF retains a lower measured time for its reported update interval: reverse sampling and proposal updates take 0.3052±0.01390.3052\pm0.0139 seconds per cycle, compared with 0.5093±0.00140.5093\pm0.0014 seconds for the complete parallel LETKF analysis update. The latter includes observation mapping, local updates, final state clipping, and GPU synchronization. AECSF's shared proposal ensemble also supports reuse of proposal-side computations and parallel score evaluation across reverse particles.

In the reported L96 comparison, LETKF has lower RMSE and CRPS; AECSF improves both metrics over EnSF and has SSR closest to one. Table 1 reports this accuracy comparison with the measured parallel LETKF runtime; Section 5.2 identifies the parallel implementation. The other methods' original measurements are retained. Appendices D.3–D.4 specify the timing intervals and parallel implementation.

AuthorsOpenReview ↗

Response to Reviewer g7e8 (2/3): Proposal and score accuracy

Q1. Direct accuracy of the adaptive proposals and guided versus likelihood-based resampling

We added two types of direct diagnostic in this revision. The first tests whether adding resampling improves proposal estimates of the reference mean and variance. The second tests whether reverse-particle guidance further improves score estimation relative to standard SMC resampling. We evaluated saved proposal ensembles at five reverse stages in ten Gaussian problems with analytic references, with four sampling repetitions per problem. The first diagnostic compares full AECSF with AECSF without resampling. The latter retains incremental-weight updates, Langevin movement, and reverse sampling, so this control examines the addition of guided resampling and its importance correction. Against the fitted tempered Gaussian reference, full AECSF has lower errors in both its mean and diagonal-variance estimates:

Proposal updateWeighted proposal-mean MSEDiagonal-variance MSE
Full AECSF0.10710.0784
AECSF without resampling0.19380.1200

The second diagnostic directly compares guided resampling with standard SMC resampling to test the additional benefit of involving reverse particles in selection. Here, standard SMC resampling selects ancestors using the tempered target weights and assigns equal weights to their descendants. Both AECSF versions retain the same reverse sampler and proposal-movement schedule. To compare score estimation itself, both proposal ensembles must be evaluated at the same locations. The score is used to advance the current reverse particles. At each saved stage, we therefore take 20 reverse-particle locations from each method's trajectory and combine them into one set of 40 evaluation points. Both proposal ensembles estimate scores at these locations. We compare these estimates with the analytic score of the noised posterior defined by the fitted prior and current observation. This avoids differences in evaluation locations while keeping the diagnostic tied to the computations needed during reverse sampling.

Resampling scheme within AECSFScore normalized mean squared error (NMSE) on the same 40 evaluation points
Reverse-particle-guided resampling0.531663
Standard SMC resampling0.557599

Guided resampling reduces score NMSE by 4.65%, with improvement in all ten problems. The tabulated errors are averaged over the five saved stages and repetitions within each problem, then over the ten problems. Appendix D.1 reports the proposal-moment and score diagnostics. Section 5.1 distinguishes the two comparisons before reporting their findings: adding resampling tests proposal approximation, while guided versus standard SMC resampling tests the additional benefit of reverse-particle information. The endpoint statistics then assess the resulting sample distribution.

Beyond these direct diagnostics, we examined the discrepancy between the final sample distribution and the fitted posterior using endpoint energy statistics. These were computed in two ways: from pooled samples and from pairs across repetitions. In the Gaussian comparison of guided and standard SMC resampling, these two statistics decrease by 31.7% and 27.6%, respectively, indicating improved final sampling quality. Beyond the Gaussian problems, the added L96 internal control tests the two resampling schemes in cycling assimilation. Reverse-particle-guided resampling reduces RMSE by 10.3% and CRPS by 7.6%, with improvement in every paired run. This provides further evidence of the benefit of guided selection for recursive filtering performance.

AuthorsOpenReview ↗

Response to Reviewer g7e8 (3/3): Sparse observations and forecast accuracy

Q2. Observation density and the diagonal Gaussian's representation of spatial correlations

Sparse observations must inform unobserved regions through spatial relationships in the forecast prior. The diagonal Gaussian in the original benchmarks describes only componentwise means and variances, limiting the dependence represented by proposal movement. To address this issue, we added a learned-prior Navier–Stokes experiment. A conditional normalizing flow represents the joint forecast density. Current observations enter through the likelihood, while AECSF retains its shared conditional-mean estimator.

We evaluate the joint prior and spatial structure in a fixed sparse configuration: 64 two-component velocity sensors on a 128×128128\times128 grid. The experiment includes four development trajectories and eight assimilation cycles. Under the same frozen prior network, we compare three proposal moves. Isotropic moves use an identity mobility matrix in the flow's generating coordinates, giving the same movement scale in each direction. Curvature preconditioning adjusts directional scales using Gaussian-reference curvature constructed from the forecast covariance and observation map. Spectral preconditioning further incorporates spectral structure estimated from additional simulated forecast fields.

MethodMean vorticity CRPS
AECSF, isotropic moves0.4939
AECSF, curvature-preconditioned moves0.4845
AECSF, spectral-preconditioned moves0.4690
LETKF (tuned separately)0.5402

All three AECSF variants have lower mean vorticity CRPS than LETKF. Spectral AECSF has the lowest CRPS and mean velocity RMSE (0.1290), while its SSR (0.8162) indicates remaining underdispersion. These results support using a spatially structured joint prior for sparse flow assimilation. Section 5.4 reports this learned-prior extension. Appendix D.6 details movement calibration: the mobility matrix acts on both drift and noise covariance, with calibration based on checks of Gaussian-reference means and directional covariances.

Q3. Forecast accuracy after cycling assimilation

We added free-forecast experiments in the paper's KS setting. At similar computational cost, AECSF has the lowest RMSE and CRPS at lead 8, showing that improved analysis accuracy can translate into better forecasts. Starting from complete analysis ensembles, we stopped assimilating subsequent observations and propagated every member with the same dynamics. The comparison uses ten paired runs and five forecast origins per run. The table reports results at a physical forecast lead of 8, where AECSF improves RMSE and CRPS against every comparator in all ten paired runs.

Method (members / sampler setting)RMSECRPSCost (s)
AECSF (160/100)0.1180 ± 0.00780.0640 ± 0.00160.5561
IEnSF-L1 (20/20)0.2536 ± 0.01770.1283 ± 0.00580.6648
EnSF (160/100)0.2816 ± 0.02530.1265 ± 0.00960.5135
EnFF-OT (160/100)0.2665 ± 0.02310.1627 ± 0.01340.4771
EnFF-F2P (160/100)0.4323 ± 0.03950.2729 ± 0.02370.5097

Parentheses give ensemble size followed by the sampling-step parameter: TT for AECSF, EnSF, and EnFF, and the number of reverse updates NtN_t for IEnSF-L1. Metrics are averaged over five origins within each run, then reported as means ± sample standard deviations across the ten paired runs. Cost includes one assimilation cycle, the complete ensemble's free forecast to lead 8, and amortized initialization; all methods were measured on the same RTX 3090.

Extending forecasts to lead 24, AECSF retains lower mean errors among these comparable-cost configurations; over leads 0–24, its time-averaged RMSE and CRPS are lower than EnSF's by 40.75% and 41.62%. In this comparison, AECSF's measured cost is 8–17% higher than the enlarged EnSF/EnFF baselines. In addition to the IEnSF-L1 configuration with 20 reverse updates in the table, we retained a higher-cost IEnSF reference with 100 updates. This reference also has larger errors at forecast lead 8. Section 5.3 and Figure 4 report the forecast curves and comparison configurations; the full horizon-8 results appear in Appendix D.5, Table 17.

Thank you again for these helpful suggestions. We have added the diagnostics and experiments, completed the parallel LETKF rerun, and clarified the method schematic, comparison settings, and timing report. We would be happy to clarify any remaining questions and continue the discussion.

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

The paper addresses a known issue of filtering methods using generative approach (notably, EnSF and EnFF), namely, the inexact guidance approximation leading to biased filtering estimates. The paper presents AECSF, which overcomes this using a Monte-Carlo estimation of the conditional score. However, since this alone is prone to particle collapse in high dimensions, AECSF considers an annealing schedule along the denoising trajectory and Langevin movement (along with appropriate importance correction) to avoid the collapse. Experiments demonstrate the benefits of adding these new components, leading to results competitive with LETKF on L96 at a fraction of the cost and much better cost-accuracy tradeoff compared to IEnSF-L1 on the Kuramoto-Sivashinsky experiment.

Critical strengths

The paper addresses a known issue in ensemble score/flow based filtering, which is the inexactness in the guidance, leading to biased filtering results. The method proposed in the paper appears to be a sound approach to address this issue. The experiments demonstrate this contribution well: e.g. the ablation studies in Figure 1-2 clearly show how each introduced component affects the overall result, and Figure 3 maps out the cost-accuracy trade-off against the EnSF-family of methods.

Critical weaknesses

Methodological positioning:

One major issue I see is positioning of the method against tempered SMC, both theoretically and experimentally. AECSF builds on a Monte-Carlo approximation of the conditional score, which is essentially a score-analogue of the MC guidance used in EnFF [Transue et al., 2026]. However, the problem with using MC guidance in EnFF is that it essentially identical to a bootstrap SMC (see Proposition 4.2 in [Transue et al., 2026]). I would like to therefore understand the framing of the work relative to tempered SMC better: are they essentially equivalent or not, and if not, exactly what additional advantage does AECSF offer? Can this difference be verified experimentally?

In addition, the experiments all use full-field observations, which I believe is much too informative to evaluate the effectiveness of the filter performance. The particle collapse issue in EnSF/EnFF makes it challenging to get it to work in sparse observation settings. Hence, it is important to understand if the proposed fix by AECSF resolves this known issue.

Finally, there are other diffusion-based DA methods that have been introduced in the last years that the paper overlooks, e.g., SDA [Rozet et al. (2023)], FlowDAS [Chen et al. (2026)], DAISI [Andrae et al. (2026)], SURGE [Wei et al. (2026)], FA-APF [Savary et al., (2026)], etc... In particular, SURGE and FA-APF are based on auxiliary particle filters and therefore I believe are quite relevant to the present work.

[Rozet et al. (2023)] Rozet, François, and Gilles Louppe. "Score-based data assimilation." Advances in Neural Information Processing Systems 36 (2023): 40521-40541.

[Chen et al. (2026)] Chen, Siyi, et al. "Flowdas: A stochastic interpolant-based framework for data assimilation." Advances in Neural Information Processing Systems 38 (2026): 103061-103102.

[Andrae et al. (2026)] Andrae, Martin, et al. "DAISI: Data Assimilation with Inverse Sampling using Stochastic Interpolants." arXiv preprint arXiv:2512.00252 (2025).

[Wei et al. (2026)] Wei, Lifu, et al. "SURGE: Approximation and Training Free Particle Filter for Diffusion Surrogate." arXiv preprint arXiv:2605.18745 (2026).

[Savary et al., (2026)] Savary, Thomas, François Rozet, and Gilles Louppe. "Training-Free Bayesian Filtering with Generative Emulators." arXiv preprint arXiv:2605.20028 (2026).

Presentation issues:

In general, I find the writing of the paper to be difficult to follow, making the reading experience slightly frustrating at parts. Some notable presentation issues include:

  • The methods introduced in Section 4.2 seem to come out of the blue. I understand that the methods introduced here are to avoid particle collapse in the computation of (12), however it would help greatly to add more detail on how each component that is introduced helps alleviate the collapse issue. At the very least, I would appreciate if there were references to where these ideas originate from, which I believe are motivated from the tempered SMC literature (e.g. the weights in (13) appear to be the weighting used at each tempering step and (16) resembles jittering). Relevant literature is listed in Section 2, but it would help to distribute these references in a more targeted fashion (see that sections 3--4 has no references at all).
  • Integration weights Wi,mW_{i,m} are introduced in line 191 before they are defined in equation (15), and it is not clear if these correspond to the same objects.
  • With most discussion deferred to the appendix, it is very difficult to understand what Section 4.4 is trying to communicate. A glaring example is line 340: "The bound separates..." --- there is no bound introduced in any discussion prior to this.
Questions for authors

I mostly understand the purpose of the methods in Section 4.2, however, one that I don't really understand is the "guided ancestor selection" (14). Could you please elaborate on how this helps over simply using the weights (13)? I am not sure what the authors mean by "Selection based only on ωi\omega_i would ignore where conditional estimates are required". Does "resampling" in Figure 2 consider this reweighing?

10 comments
AuthorsOpenReview ↗

Response to Reviewer WryB (1/4): Relationship to tempered SMC

We sincerely thank you for your careful reading of our manuscript and for your insightful questions and valuable suggestions. Your comments are very helpful in clarifying the method and improving the manuscript. We agree with your assessment that AECSF is closely related to SMC. AECSF draws on tempered SMC to update its proposals. These shared proposals estimate the conditional means needed for reverse sampling, while the current reverse particles help determine how the proposals are allocated. To study this interaction, we added comparisons with independent SMC methods and controls within AECSF, examining the overall construction and the role of guided selection, respectively. We first explain the methodological relationship and comparison results, then address sparse observations and the presentation of the method.

W1. Relationship to tempered SMC, the bootstrap-filter connection in EnFF, and additional value

AECSF has a direct methodological connection to tempered SMC. Understanding both the connection and the distinction begins with the computational task assigned to the proposals. A noisy reverse particle is compatible with multiple clean states, which differ in their compatibility with the observation. The conditional Tweedie identity turns score computation into conditional-mean estimation: averaging clean states conditioned on both the current noisy particle and the observation. Monte Carlo is a natural way to compute this mean. The difficulty is that different reverse particles require different conditional averages, and those averages change throughout denoising. A fixed proposal ensemble may not adequately cover the regions needed for these computations. Thus, beyond weighting existing proposals, the estimator needs a continuing supply of proposals suited to the current computation. Likelihood weighting, resampling, and movement in tempered SMC provide a basis for progressively adapting those proposals.

In the independent tempered SMC configurations compared here, the particles directly form the posterior approximation. AECSF instead uses two particle ensembles: shared proposals estimate conditional means, and reverse particles use these estimates to advance reverse sampling and form the analysis ensemble at the endpoint. This distinction means that proposal allocation also needs to account for the current reverse particles. A tempered target weight describes a proposal's importance to the target distribution. Yet the same proposal can contribute very differently to the conditional means needed by different reverse particles. Gaussian compatibility supplies this latter information. Guided selection therefore combines target weights with this compatibility to allocate shared proposals, with importance correction accounting for the changed selection probabilities. The updated proposals then provide the scores needed for the next reverse-sampling step. Q1 explains the allocation rule in more detail.

The tempered reference combines the forecast prior with LρiL^{\rho_i}, allowing proposals to adapt gradually to observations. When estimating conditional means, L1−ρiL^{1-\rho_i} supplies the likelihood not yet included in this reference, alongside Gaussian compatibility with the current reverse particle.

The connection between EnFF and bootstrap filtering that you identified is explicit in the special case of fixed proposals. Suppose that proposal locations remain unchanged, with no resampling or movement, and only incremental likelihood weights are accumulated. Let umu_m denote the initial proposal weights at tempering level zero, L(zm)L(z_m) the current observation likelihood at proposal zmz_m, and ρi\rho_i the current tempering level. Multiplying the accumulated weight Wi,mW_{i,m} by the remaining likelihood factor gives

Wi,mL(zm)1−ρi∝umL(zm)ρiL(zm)1−ρi=umL(zm).W_{i,m}L(z_m)^{1-\rho_i}\propto u_mL(z_m)^{\rho_i}L(z_m)^{1-\rho_i}=u_mL(z_m).

The two likelihood powers sum to one, recovering the full likelihood; Appendix D.4 gives this derivation. This special case therefore yields the score of a Gaussian-smoothed, likelihood-weighted empirical posterior. With compatible initialization, exact scores, and exact reverse dynamics, terminal sampling can be understood as three operations. It selects a center according to the likelihood-reweighted weights, multiplies it by the terminal signal coefficient, and adds the retained terminal noise. This explains the connection to bootstrap filtering while retaining the effect of finite terminal noise. Full AECSF updates proposal locations and weights throughout denoising and lets the current reverse particles influence proposal allocation. It draws on SMC to adapt proposals for the changing conditional-mean computations required by reverse sampling.

AuthorsOpenReview ↗

Response to Reviewer WryB (2/4): SMC comparisons and guided selection

W1 (continued). Experimental comparisons

To test the practical effect, we designed two types of comparison in the paper's 10,000-dimensional L96 setting. The independent SMC comparisons examine the overall construction; the controls within AECSF examine guided selection and its correction. The independent SMC comparison includes two configurations. Factorized tempered SMC uses the product structure formed by a diagonal forecast prior and coordinatewise likelihoods. The SMC variant with transport and marginal moment matching changes only how the output ensemble is constructed, matching the weighted mean and variance of each state component.

The internal control keeps AECSF's reverse sampler and proposal moves unchanged and compares only the ancestor-selection rules. Standard SMC resampling selects ancestors using tempered target weights and assigns equal weights to the resampled descendants. Reverse-particle-guided resampling incorporates the current reverse particles into selection and applies the corresponding importance correction.

MethodRMSECRPS
Factorized tempered SMC2.7330 ± 0.03771.7974 ± 0.0340
SMC + transport and marginal moment matching2.6054 ± 0.02651.6862 ± 0.0240
AECSF, standard SMC resampling0.1708 ± 0.00540.0840 ± 0.0010
AECSF, reverse-particle-guided resampling0.1531 ± 0.00440.0776 ± 0.0008

Entries are means ± sample SDs over ten paired runs using the paper's evaluation window. All methods output 20 members: SMC uses 320 internal particles and AECSF 20 shared proposals.

AECSF with standard SMC resampling already substantially outperforms both independent SMC configurations, supporting the practical value of the overall construction on this benchmark. With the reverse sampler and proposal moves held fixed, guided selection and its correction yield a further 10.3% reduction in RMSE and 7.6% in CRPS. Both improvements occur in every paired run. This internal control supports the additional benefit of allocating proposals according to the current reverse particles.

Table 1 includes both independent SMC configurations and AECSF with standard SMC resampling. Appendix D.4, Table 8 retains each complete paired comparison. Sections 4.1–4.2 explain how SMC operations support conditional-mean estimation; Section 5.2 distinguishes the overall comparison from the internal test of guided selection.

Q1. Why guided ancestor selection, and what does Figure 2's resampling include?

The target weights in Eq. (13) describe a proposal's importance to the tempered reference distribution. Selecting proposals solely from these weights does not yet account for the current reverse-particle locations or the conditional means that must be estimated there. Equation (14) incorporates this information, which is what the original phrase “where conditional estimates are required” means. The allocation is computed in three steps:

  1. For each reverse particle, multiply the proposals' target weights by their Gaussian compatibility, which measures how well a noised proposal matches that reverse particle's location.
  2. Normalize the allocation weights separately for each reverse particle.
  3. Average these allocations across reverse particles so that all reverse particles contribute to the selection probabilities of the shared proposals.

Ancestor importance correction accounts for selecting ancestors with these probabilities instead of directly using the target weights. The remaining likelihood factor enters the conditional-mean coefficients, not this ancestor-allocation rule. The resampling in both full AECSF and the original “resampling-only” variant includes reverse-particle-guided ancestor selection and its importance correction. Revised Figure 2 names the latter “No Langevin moves”. The labels below state which operations are retained; all variants keep incremental likelihood-weight updates and reverse sampling:

Label in revised Figure 2Proposal update
No Langevin movesGuided ancestor selection and importance correction; no movement.
No resamplingLangevin movement; no ancestor selection.
Fixed proposal positionsNeither resampling nor movement; weights still update.
Standard SMC resampling + LangevinTempered-weight ancestor selection, equal-weight descendants, and Langevin movement.

The final control changes the ancestor-selection rule, whereas the preceding variants remove operations. Section 5.2 and Appendix D.4 define these controls; the shorter Figure 2 caption directs readers there.

AuthorsOpenReview ↗

Response to Reviewer WryB (3/4): Sparse observations and related work

W2. Sparse observations and particle collapse

We added an experiment with a learned-prior extension under sparse Navier–Stokes observations. Sparse observations pose a particular difficulty: measurements at a few locations must help infer a much larger unobserved field. The forecast prior therefore needs to represent spatial relationships between measured quantities and the rest of the state. For this purpose, we use a learned prior over the joint flow field while retaining AECSF's shared conditional-mean estimator. The conditional normalizing flow is trained on simulated forecast ensembles and independent target fields, then frozen during assimilation; the current observation still enters through the likelihood.

The experiment uses a 128×128128\times128 Navier–Stokes system with viscosity 0.02, 64 two-component velocity sensors, and 32 analysis members. Evaluation comprises four independently generated development trajectories, with eight assimilation cycles per trajectory. Under the same prior network, we compare three types of proposal movement. Each variant still uses its own evolving forecast context; LETKF parameters were selected on separate tuning trajectories and then held fixed. In all three variants, the mobility matrix acts on both the drift and noise covariance.

  • Isotropic moves: Use an identity mobility matrix in the flow's generating coordinates, giving the same movement scale in every direction.
  • Curvature-preconditioned moves: Adjust directional movement scales using Gaussian-reference curvature constructed from the forecast covariance and observation map.
  • Calibrated spectral moves: Further incorporate spectral structure into the preconditioning; this structure is estimated from additional simulated forecast fields.
MethodVelocity RMSEVorticity CRPSVorticity SSR
AECSF, isotropic moves0.1339 ± 0.01930.4939 ± 0.05980.7356 ± 0.0548
AECSF, curvature-preconditioned moves0.1358 ± 0.03640.4845 ± 0.11630.7878 ± 0.1419
AECSF, calibrated spectral moves0.1290 ± 0.02360.4690 ± 0.07930.8162 ± 0.0896
LETKF0.1586 ± 0.02560.5402 ± 0.04531.4098 ± 0.1275

Entries are means ± sample standard deviations of four trajectory-level summaries. RMSE and CRPS are averaged over eight cycles; the spread–skill ratio (SSR) pools second moments within each trajectory, following the manuscript definition. Calibrated spectral AECSF reduces mean RMSE and CRPS by 18.7% and 13.2% relative to LETKF. Its trajectory-averaged CRPS is lower in all four paired comparisons. SSR also increases from 0.7356 with isotropic moves to 0.8162, closer to one. It nevertheless remains below one, indicating underdispersion. Section 5.4 presents this experiment as a learned-prior extension and explains how decoded proposals enter the shared conditional-mean estimator. Appendix D.6 gives the prior interface and movement calibration based on Gaussian-reference means and directional covariances.

W3. SDA, FlowDAS, DAISI, SURGE, and FA-APF

Section 2 compares these works by the source of their generative distribution and where observations enter the algorithm. SDA was already cited; the revision makes its relationship to AECSF explicit alongside the other four methods.

MethodDiscussion in the revised related work
SDALearns trajectory scores.
FlowDASLearns state transitions through stochastic interpolants and uses Monte Carlo likelihood guidance.
DAISIConnects a learned prior to the forecast ensemble through inverse sampling.
SURGEWeights and resamples guided diffusion paths with a Girsanov correction.
FA-APFUses predictive-likelihood weights to select ancestors, then generates descendants through an approximate observation-conditioned sampler.

SURGE and FA-APF are discussed together as particle methods using pretrained generative emulators, to explain their connection to this work. The original AECSF construction obtains its analysis score from the current forecast ensemble and likelihood. Selection and importance correction operate on shared proposals within one analysis. These proposals supply conditional-mean estimates to the reverse particles, which form the final output. This makes the distinction between proposal adaptation and output samples explicit.

Comments on the paper

3 comments

Dear authors and reviewers,

The author-reviewer discussion window will close tomorrow on November 18.

@authors: If not done already, please respond to the reviewers' comments before the discussion window closes. I would advise short and concise responses over lengthy explanations.

@reviewers: Please check the authors' responses and provide any additional comments before the discussion window closes. Consider updating your initial review if necessary, and in particular your score if your opinion has changed. You may increase your score if you feel the authors' responses warrant it, maintain your score if your opinion remains the same, or even decrease your score if you feel the other reviews or the authors' responses reveal new shortcomings. At the very least, please acknowledge the authors' responses.

Thank you for your attention and cooperation.

The AC

AuthorsOpenReview ↗

General response to all reviewers and the AC

We thank all reviewers and the Area Chair for their time, careful evaluation, and constructive feedback. These comments have been very helpful in improving our manuscript. The main updates are:

  1. Method and theory. We clarified how shared proposals yield conditional scores, and expanded the explanations of the connections to tempered SMC and KDE conditioning, as well as the error analysis.
  2. Additional comparisons and diagnostics. We added independent tempered SMC baselines and paired Lorenz–96 comparisons of Binder, enhanced Binder, and AECSF, with the enhanced baseline using the original observation likelihood. We also evaluated proposal and score-estimation accuracy and the benefit of reverse-particle-guided selection.
  3. Parallel LETKF benchmark. We remeasured LETKF with parallel GPU execution and updated the efficiency comparison and timing descriptions.
  4. Forecast evaluation. We added KS forecast comparisons at multiple lead times, using complete analysis ensembles and comparable computational budgets.
  5. Sparse observations and learned priors. We extended the learned-prior Navier–Stokes study to 50 cycles on four paired development trajectories and added forecast-only, EnSF, and EnFF baselines.
  6. Workflow schematic. We added and refined a workflow schematic to distinguish the forecast–analysis and denoising cycles and show more clearly how shared proposals and reverse particles interact.

With the discussion window closing soon, we welcome further discussion and would be happy to clarify any remaining questions. Thank you again for your time and helpful suggestions.

AuthorsOpenReview ↗

Follow-up to our general response

Since our previous summary, we have revised the manuscript to make the method easier to follow and the purpose of each experiment clearer. The most substantial changes are in Section 4 and Appendix D.6. Section 4 now explains how AECSF uses either a fitted Gaussian prior or a learned forecast prior, how each guides proposal moves, and how both lead to the same conditional-score estimator. Appendix D.6 follows the learned prior from its construction through the analysis update, with the full procedure summarized in the new Algorithm 3.

We also revised the openings of the experimental sections to explain what each comparison tests. New figures for the sparse Navier–Stokes experiment show how the reconstructed vorticity fields evolve and where errors persist. The main revisions are highlighted in blue in the manuscript.

We sincerely thank the reviewers and the Area Chair for their time and for the thoughtful feedback and constructive suggestions they have shared throughout the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.