acceptodds
Under review as a conference paper at ICLR 2027

Generative Model Proposal based Particle Filtering for Data Assimilation

Abstract

Data assimilation (DA) models state dynamics conditioned on sequential observations, with wide-ranging scientific applications. In filtering, the goal is to model the posterior over the current state given all observations so far. Classical solutions make simplifying distributional or functional assumptions, e.g., linear-Gaussian systems, which can be inaccurate. In principle, particle filters remove these assumptions, yet often collapse in high dimensions. Recent generative DA approaches learn conditional state transitions, but typically incorporate observations through approximate guidance, so errors can accumulate over long horizons. We introduce Flow Proposal Particle Filters (FPPF), which learn a conditional generative-model-based proposal approximating the variance-minimizing optimal proposal. Conditioning on observations steers particles toward high-likelihood regions before weighting, reducing weight variance and delaying degeneracy. Since our proposal admits tractable likelihood evaluation, FPPF computes accurate importance weights and retains a Bayesian update. We further introduce L-FPPF, which uses localization to scale FPPF to high-dimensional systems where global particle filters collapse. Experiments on chaotic dynamical systems and multimodal observations of severe convective storms show that FPPF improves accuracy and calibration over several statistical baselines and other generative DA methods, especially in strongly nonlinear settings, and remains stable over long horizons.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
data assimilation, flow matching, particle filtering, bayesian filtering, generative modeling, weather

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_E2hNOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a method for data assimilation (DA) using generative models. Specifically, they propose Flow Proposal Particle Filters (FPPF), which learns the optimal proposal using a flow model. Recent methods have used flow-based generative models to perform filtering in high-dimensional settings by guiding towards observations during inference. The guidance methods used in these approaches are approximative, incurring an irreducible bias. Further, methods that try to sample from the optimal proposal either ignore the necessary weight updates or approximate them similarly to guidance. FPPF instead uses the learned flow model to propagate the likelihood together with the sample, avoiding these biases. They also consider a localized variant which avoids particle collapse in high-dimensions. In a range of experiments across different settings, the method performs decent, showing promising results.

Critical strengths

Despite leaning towards rejection, the work has some merit. In particular, the localized version is interesting for avoiding weight collapse and the noisy trace estimator. Also, the cross-modal SEVIR experiments with learnable likelihoods are interesting and goes beyond what most generative methods for DA attempts.

Critical weaknesses

My support of rejecting this paper is based on two main concerns: lack of novelty and inconclusive experiments.

Lack of novelty.

My first main concern is that the paper oversells its novelty.

The authors are not the first ones to use conditional generative models in fully-adapted particle filters. The Normalizing Flow-based filters (NF-DPF) of [2, 3] learn to sample from p(xt∣xt−1,yt)p(x_t|x_{t-1},y_t) using conditional normalizing flows and calculate the weights using this. Further, FPPF replaces the cheap tractable log-likelihoods of NF-DPF with Hutchinson trace estimators that are known to be high variance in high-dimensions. The authors cite [2,3], but don't compare to them and write in A.1 that:

To our knowledge, FPPF is the first to instantiate the proposal with a conditional generative model carrying no distributional assumption while retaining tractable density evaluation

which is overselling its novelty.

The APF of [1] is motivated by the existence of high-quality pretrained flow-based generative models that would be too expensive to retrain on your own. With observations during training, this benefit is lost, and we have to retrain for every new observation. The authors write that:

Given a pretrained generative forecaster, we can freeze its backbone and train only a lightweight adapter

but it’s not something they test or elaborate on.

Inconclusive experiments

My second main concern is that the experiments don't demonstrate the significance of the paper.

Firstly, FPPF has an unfair advantage to its baselines since it sees observations during training. This is not the problem the baselines were designed for, and it's thus expected that a method train directly on observations would outperform one which only sees them during inference. The authors don't compare to the very related normalizing flow-based filter [3], or to other baselines with access to observations during training.

Despite this advantage, the results show marginal gains over baselines. On L96, the inference-time guided DAISI method is on par with LPPF, despite the main claim of approximative guidance being a big problem. On both SQG and SEVIR, FPPF have similar performance to the localized BPF, suggesting that the performance comes from the localization alone rather than FPPF.

The SQG experiment seems rushed, ignoring the sparser, more nonlinear settings considered in [4] as well as missing experimental details such as number of ensemble members. The DAISI method presented in [4] performs poorly despite being tuned on that problem, raising the question if the authors implemented it correctly.

[1] Savary, T., Rozet, F., and Louppe, G. Training-Free Bayesian Filtering with Generative Emulators. ICML 2026.

[2] Chen, Xiongjie, Hao Wen, and Yunpeng Li. "Differentiable particle filters through conditional normalizing flow." 2021 IEEE 24th International Conference on Information Fusion (FUSION). IEEE, 2021.

[3] Chen, Xiongjie, and Yunpeng Li. "Normalizing flow-based differentiable particle filters." IEEE Transactions on Signal Processing 73 (2024): 493-507.

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

Questions for authors
  • Provide a comparison to a normalizing-flow based filter and a motivation of FPPFs novelty over those methods.
  • Show that LFPPF actually performs better than LBPF to rule out the localization being the sole reason it performs well.
  • Expand the SQG experiments to cover the more difficult settings.
  • Verify that DAISI is implemented correctly according to the hyperparameters in that paper.
  • Verify the claim of finetuning a pretrained generative forecaster or drop it.
8 comments

Thank you for the detailed assessment. We added a direct NF-DPF [3] comparison, expanded SQG to the four observation settings of DAISI [4], and clarified the novelty claim and experimental scope. Changes are in blue in the revised PDF.

W1 / Q1. Novelty and comparison with NF-DPF.

We revised the claim in Appendix A to acknowledge that NF-DPF [2, 3] already uses conditional generative proposals with tractable importance weights. FPPF differs in how the proposal is learned and in how it scales. NF-DPF proposes conditional normalizing flow proposals inside a differentiable particle filter, jointly learning transition, observation and proposal models by optimizing sequence likelihood or filtering accuracy. In contrast, we propose a more modular framework in which known forecast and observation models are used directly, and the proposal is trained separately by rectified flow regression on historical state-observation pairs. Training therefore needs no filtering rollouts and no differentiation through weighting or resampling. We found it stable in practice, which let us train large spatial proposal networks when necessary.

Moreover, L-FPPF introduces localized proposal-density corrections, enabling experiments at substantially higher state dimensions. NF-DPF evaluates systems with up to 100 state variables, whereas our experiments extend to 1,000 dimensions in Lorenz–96, 8,192 in SQG, and 16,384 in SEVIR.

New NF-DPF comparison. To compare the two proposal learning approaches directly, we implemented the NF-DPF baseline using the published code from [3]. For fair comparison, we replace the learned dynamics and observation likelihood with the known distributions, and keep our Lorenz–96 evaluation settings. While NF-DBF’s learned proposal improved on BPF, FPPF achieves lower mean RMSE and CRPS than NF-DPF across all six dimensions and both observation operators, as shown below. Full results with standard deviations are provided in Appendix Tables 1-2, and we have updated the plots in Figure 2 to include NF-DPF. Entries report RMSE / CRPS.

dxd_xBPFNF-DPFFPPF
h(x)=arctan⁡(x)h(x) = \arctan(x)
50.303 / 0.1950.300 / 0.1950.189 / 0.115
100.357 / 0.2120.351 / 0.2100.287 / 0.165
150.364 / 0.2120.352 / 0.2080.217 / 0.122
200.419 / 0.2330.391 / 0.2210.238 / 0.132
250.480 / 0.2620.446 / 0.2450.235 / 0.132
500.743 / 0.4390.659 / 0.3810.323 / 0.183
h(x)=min⁡(x4,10)h(x) = \min(x^4, 10)
50.279 / 0.2050.282 / 0.2050.201 / 0.125
100.411 / 0.2500.412 / 0.2490.321 / 0.191
150.404 / 0.2420.410 / 0.2460.237 / 0.134
200.591 / 0.3320.528 / 0.2960.382 / 0.223
250.803 / 0.4560.715 / 0.4010.254 / 0.142
501.942 / 1.2131.687 / 1.0300.519 / 0.289

NF-DPF improves on BPF mainly at dx≥20d_x \geq 20 and is close to it below. FPPF has lower RMSE and CRPS than NF-DPF in all 12 settings, and the gap is largest at dx=50d_x = 50 (RMSE 0.323 vs 0.659 under arctan⁡\arctan, 0.519 vs 1.687 under min⁡(x4,10)\min(x^4, 10)). Full results with standard deviations are in Appendix Tables 1 and 2, and Figure 1 now includes NF-DPF.

Training-time observation access. These results also address the concern that FPPF's gains arise simply because it sees observations during training while baselines do not. The NF-DPF proposal likewise uses observations during training as do the NASMC and InfNN baselines already in the submitted appendix. We will discuss their training approaches explicitly in the main experimental section.

W2 / Q3-Q4. Expanded SQG experiments and DAISI reproduction.

As requested, we expanded SQG to all four observation settings of DAISI [4]: Noisy (h(x)=xh(x) = x, σobs=5\sigma_{obs} = 5, 25% observed), Sparse (xx, 1, 5%), Saturating (arctan⁡x\arctan x, 0.01, 25%) and Multimodal ((x/7)2(x/7)^2, 1, 25%). For DAISI we use the authors' released SQG-64 checkpoint and sampling code, with 100 inverse and 100 forward Euler steps as in their paper. Our DAISI results closely match the published values on Sparse, Saturating and Multimodal (RMSE / CRPS):

DAISINoisySparseSaturatingMultimodal
Reported in [4]2.46 / 1.323.40 / 1.732.96 / 1.543.71 / 1.81
Ours4.509 / 2.6993.386 / 1.7232.804 / 1.4553.688 / 1.812

On Noisy, our DAISI error remains about twice the reported value, despite using the released checkpoint and sampler and rechecking all settings.

Expanded results (RMSE / CRPS):

SettingL-FPPF (N=200N{=}200)LBPF (N=200N{=}200)DAISI (N=20N{=}20)LETKF (N=50N{=}50)
Noisy3.259 / 1.7973.271 / 1.8124.509 / 2.6992.86 / 1.53
Sparse3.135 / 1.6544.890 / 2.6043.386 / 1.7233.694 / 1.922
Saturating1.758 / 0.9192.061 / 1.0532.804 / 1.4554.588 / 2.636
Multimodal3.044 / 1.6086.556 / 3.5793.688 / 1.8123.023 / 1.625

We tune the localization radius separately for L-FPPF, LBPF and LETKF in each setting. The selected configurations, ensemble sizes and remaining details are added to the appendix, and Section 5.3 now includes the full results.

W2 / Q2. Gains over LBPF and DAISI.

The expanded SQG experiments show substantial improvements over LBPF and DAISI in challenging observation settings. LBPF is competitive on Noisy, but L-FPPF lowers RMSE relative to LBPF by 35.9% on Sparse, 14.7% on Saturating and 53.6% on Multimodal. Since L-FPPF and LBPF share the same localization, these gains come from the learned proposal and not from localization alone. LETKF remains best on Noisy, where its linear-Gaussian assumptions hold, and ties L-FPPF on Multimodal (RMSE 3.023 vs 3.044, CRPS 1.625 vs 1.608).

On high-dimensional Lorenz-96, DAISI is competitive in RMSE under h(x)=arctan⁡(x)h(x) = \arctan(x), but L-FPPF has 27% to 30% lower CRPS at every dxd_x (0.156 to 0.162 vs 0.222 to 0.224). Under h(x)=min⁡(x4,10)h(x) = \min(x^4, 10), L-FPPF lowers RMSE by 41% to 47% and CRPS by about half relative to DAISI at every dxd_x from 50 to 1,000 (Table 4).

Both results highlight that our learned proposal performs well under sparse, linear, and highly nonlinear observation operators alike, while inference-time guidance methods struggle when the observation operator is highly nonlinear.

Q5. Adapting a pretrained generative forecaster.

Our SEVIR experiment already does this, where the generative transition backbone is frozen, and the observation conditioning is trained as an adapter to form the proposal. We will expand Section 5.4 and Appendix C.2 to state which components are frozen and which are trained, with parameter counts. Following ControlNet (Zhang et al., 2023), the adapter freezes the 30.08M-parameter forecaster and adds 200,960 trainable parameters (0.67%), which inject the observations into intermediate flow network layers.

I thank the authors for answering my questions and running new experiments. Below are some more thoughts.

NF-DPF: I thank the authors for implementing this and for clarifying the differences to FPPF. I incorrectly thought that NF-DPF (and related models) used a MLE-based training objective to get a similar training procedure to FPPF. Thanks for clarifying this, the experiments show a clear advantage of FPPF over NF-DPF as trained here.

However, I did find this work [1], which in Sec. 3.2. Flow-based particle filtering presents a normalizing flow-based filter that is closer to what I thought NF-DPF was. This is the comparison I had in mind. I recognise this is a late request and that [1] is fairly concurrent (Apr 2026), but it should be easy to run with the NF-DPF setup you have. Train your existing RealNVP by negative log-likelihood on the same (xt−1,ot,xt)(x_{t−1}, o_t, x_t) tuples used for FPPF, and plug it into the same SIR filter. Please also add some notes on why FPPF would be preferable to the method in [1].

Still, I don't think this fully resolves my point about FPPF having an unfair advantage to its baselines since it sees observations during training. This point should be made even more clear in the main paper.

SQG: I thank the authors for including the new SQG results. The new results look promising, and show a clear gain of LFPPF over LBPF. Some more thoughts on this

  • Could you provide some more visualization of the results? From numbers alone it's hard to say why LFPPF clearly beats LBPF on sparse and multimodal but performs similarly on saturating and noisy. Are the samples realistic or do they have artefacts?
  • LFPPF and LBPF uses a much higher NN than LETKF and DAISI (200 vs 50/20), which is not reflected in the main paper. The matched inference-cost is a good motivation, but should be stated explicitely. Make sure to report the NN-adjusted fair CRPS to avoid biased comparisons.
  • I'm also curious how well LFPPF would work with a smaller N=20N=20 and how this affects the relative order.
  • Figure 3 compares LFPPF with DAISI on noisy, where you were not able to reproduce the original results. Unless you believe that the original DAISI paper has a flaw in its noisy experiments, this comparison might be misleading.
  • Is there anything different in the configuration between noisy and the other settings that can explain the DAISI lack of performance (new appendix C.3 says DPS/MMPS guidance while old version says MMPS)?

Pre-trained forecaster: This resolves my concern, and I ask only that you clarify this in the main paper.

[1] Cui, Tiangang, et al. "Amortized Filtering and Smoothing with Conditional Normalizing Flows" arXiv preprint arXiv:2604.07169v1 (2026). https://arxiv.org/abs/2604.07169v1

Some of my concerns were addressed during the rebuttal and I'm raising my score to a 2 pending the remaining answers.

Reviewer_CZb1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Flow Matching Proposal-based Particle Filtering (FPPF), which learns an observation-conditioned proposal distribution using Flow Matching and incorporates it into particle filtering through importance weighting. The authors further introduce a localized version (L-FPPF) for high-dimensional data assimilation. Experiments on Lorenz-63, Lorenz-96, SQG, and SEVIR demonstrate promising results, particularly under nonlinear observation settings.

Critical strengths
  1. Well-motivated approach. Combining Flow Matching with particle filtering is a reasonable and principled approach. The tractable proposal density enables importance-weight correction while allowing flexible, observation-conditioned sampling.
  2. Comprehensive experiments. The paper evaluates the proposed method across several dynamical systems, covering different dimensions, nonlinear observation operators, and both synthetic and real-world data. The authors compare against traditional particle filters, LETKF, learned-proposal methods, and generative data assimilation approaches. The results demonstrate clear improvements in several nonlinear observation settings.
Critical weaknesses
  1. Potential training–inference mismatch. During training, the proposal is learned from true state trajectories, where the previous state and current observation are statistically consistent. However, during inference, the previous state is an imperfect analysis particle that may be inconsistent with the current observation. It is unclear how robust the learned proposal is under such conditions, particularly during long-term cycling.
  2. Limited improvements on more complex systems. Although the results on Lorenz-96 are promising, the improvements are relatively modest on SQG and SEVIR. In particular, L-FPPF performs similarly to LBPF on SQG and is outperformed by LETKF. Considering the additional computational cost of flow sampling and density evaluation, the practical advantage is not fully established. Also, I am afraid of the fairness of the LETKF baseline on SEVIR. L-FPPF uses a learned observation likelihood for satellite infrared observations, but the paper does not clearly explain how the observation operator and observation error are constructed. This makes it difficult to assess whether the comparison is conducted under comparable observation-model assumptions.
Questions for authors
  1. How does the learned proposal perform when the previous analysis particle is substantially inconsistent with the current observation? Could the authors evaluate its robustness under different levels of previous-state error?
  2. Could the authors provide an accuracy–cost comparison between L-FPPF, LBPF, and LETKF, particularly for the SQG and SEVIR experiments?
  3. Given that EnKF/ETKF-informed proposals have been explored in particle filtering, have the authors considered comparing against a Kalman-informed proposal combined with importance weighting and resampling? This could help clarify the advantages of the learned flow proposal over simpler observation-informed proposals.
  4. For the SEVIR experiment, how are the observation operator and observation error covariance defined for the LETKF baseline? Are they based on observation-model information comparable to that used by L-FPPF?
5 comments

Thank you for the review and the specific questions. Some requests overlap with other reviewers; for those we summarize here and point to the shared results. Changes are in blue in the revised PDF.

W1 / Q1. Robustness when the previous particle is inconsistent with the observation.

It is true that FPPF learns its proposal from simulated trajectories with process noise, but during filtering it could condition on previous particles that may differ more substantially from the true state as errors accumulate over time. However, two parts of the method limit this effect:

  • The SIR weight p(ot∣xt) p(xt∣xt−1)/qϕ(xt∣xt−1,ot)p(o_t \mid x_t)\,p(x_t \mid x_{t-1}) / q_\phi(x_t \mid x_{t-1}, o_t) inherently corrects proposal error. A poor proposal for an inconsistent particle lowers its weight and the ESS.
  • During training we mask coordinates of xt−1x_{t-1} and drop oto_t (Appendix C.2), so the network learns to rely on oto_t when xt−1x_{t-1} is uninformative, instead of trusting xt−1x_{t-1} alone.

We also check robustness empirically, and add the experiments to the Appendix Section A.1.3. On Lorenz-96 (dx=20d_x = 20, h=arctan⁡h = \arctan, N=5000N = 5000), we run FPPF for 50 updates and then shift every particle by the same random vector, preserving the ensemble spread. The shift magnitude is αr0\alpha r_0, where α∈{0,0.5,1,2}\alpha \in \{0, 0.5, 1, 2\} and r0=0.233r_0 = 0.233 is the reference RMSE, so the perturbation is scaled relative to the typical filtering error. Starting from the same perturbed particles and weights, we run FPPF and BPF with unchanged observations. Results are averaged over 32 test trajectories and three shift directions per trajectory. The table below reports RMSE after the first filtering update; please see the appendix figure for the following 40 updates.

Added shiftPrevious mean RMSE (after shift)FPPF next RMSEBPF next RMSE
0×0.2440.2420.242
0.5×0.2740.2660.264
1×0.3440.3230.319
2×0.5340.4920.480

Both filters experience nearly identical increases in RMSE after the first filtering update as the shift grows. This suggests that the learned proposal is not substantially more sensitive to errors in the previous particles than the standard transition proposal used by BPF. Over the following 40 updates after a 1× shift, FPPF maintains substantially lower filtering error than BPF, showing that its learned proposal remains effective even when the previous particles are inaccurate.

W2 / Q2. Gains over localized baselines.

We ran the cost comparison and a matched wall-clock comparison on SQG for Reviewer TU32, and full tables are in that response. In brief, one assimilation step on an A100 takes 5.74 s for L-FPPF (N=200N = 200), 3.85 s for LBPF (N=200N = 200) and 7.86 s for LETKF (N=50N = 50). At equal wall-clock time (LBPF N=300N = 300, LETKF N=33N = 33), L-FPPF has the lowest RMSE and CRPS in three of the four SQG observation settings of DAISI (Andrae et al., 2026), which we added at the request of Reviewer E2hN. Against LBPF it lowers RMSE by 35% (Sparse), 13% (Saturating) and 52% (Multimodal).

We agree that in the SQG setting of the submission (linear observations with Gaussian noise), LETKF is best and L-FPPF and LBPF are tied. That setting matches the assumptions of LETKF, and the observations are weakly informative, so a learned proposal has little to add.

Q3. Comparison with a Kalman-informed proposal.

We agree this is a valuable comparison, and add a corresponding experiment. For each particle, we linearize hh around its forecast mean f(xt−1(i))f(x_{t-1}^{(i)}) and apply the Kalman update with transition covariance QQ and observation covariance RR. This gives a Gaussian approximation to the same target q⋆q^\star (local linearization, Doucet et al., 2000). EnKF-informed variants such as the weighted EnKF (Papadakis et al., 2010) use the same update with an ensemble covariance. Particles are weighted with the full SIR ratio and resampled. We test it in the Lorenz-96, h=arctan⁡h = \arctan setting, RMSE / CRPS reported below.

dxd_xBPFKalman-proposal PFFPPF
50.303 / 0.1950.298 / 0.1930.189 / 0.115
100.357 / 0.2120.337 / 0.2040.287 / 0.165
150.364 / 0.2120.336 / 0.2020.217 / 0.122
200.419 / 0.2330.372 / 0.2140.238 / 0.132
250.480 / 0.2620.407 / 0.2270.235 / 0.132
500.743 / 0.4390.561 / 0.3040.323 / 0.183

The Kalman proposal improves on the bootstrap proposal, and its gain grows with dxd_x in the global filter as weights degenerate. However, FPPF maintains the lowest error at every dimension.

Q4. LETKF observation model on SEVIR.

LETKF uses the same learned transition p^ψ\hat p_\psi as L-FPPF and LBPF for its forecast. Its observation model is trained on the same paired VIL-infrared data with the same architecture as the learned likelihood p^η\hat p_\eta used by L-FPPF and LBPF, except for L-FPPF and LBPF we learn a flow matching velocity network, while for LETKF the network has a head that outputs a Gaussian mean and variance, which serve as LETKF's observation operator h(x)h(x) and error covariance RR. We will add these details to the paper.

References: Papadakis, N., Mémin, E., Cuzol, A., and Gengembre, N. Data assimilation with the weighted ensemble Kalman filter. Tellus A, 62(5):673-697, 2010.

I thank the authors for their detailed responses and additional experiments. While several of my concerns have been addressed, I still have two reservations:

  1. Training–inference mismatch. The robustness experiment is useful, but applying a common shift to all particles does not fully represent the errors encountered in practical data assimilation. The potential distribution mismatch between training and long-term cycling remains insufficiently investigated.
  2. Experimental fairness and reliability. The ensemble sizes differ substantially across methods in the SQG and Lorenz-96 experiments. Some performance gaps are also unexpectedly large based on my experience. I will maintain my original score.
Reviewer_NqrpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Flow Proposal Particle Filter, which uses observation conditioned rectified flow to learn to approximate p(xt∣xt−1,otp(x_t\mid x_{t-1},o_t. The authors also use SIR importance weighting to correct the proposal error. To mitigate the particle degeneracy problem in high-dimensional systems, the author further introduce the localized variant L-FPPF. These method are demonstrated with competitive accuracy across chaotic dynamical systems and weather data assimilation.

Critical strengths
  1. The empirical evaluation is broad and closely aligned with the paper's theoretical motivation. The experiments are largely designed around the specific failure modes discussed in the paper, such as particle degeneracy, nonlinear observation updates, and multimodality.
  2. The method provides a cline integration of flexible generative modeling with classical particle filtering. The design combines the expressive power for generative models with principled Bayesian correction mechanism.
Critical weaknesses
  1. The computational cost is not sufficiently characterized. Although the paper acknowledges the inference cost, it does not provide the wall-clock runtime (and also the training time) comparison against other mother such as LETKF, LBPF, DAISI, or FA-APF. It is hard to understand the trade off between accuracy and the additional computational overhead.

  2. The learned proposal qϕ(xt∣xt−1,ot)q_\phi(x_t\mid x_{t-1},o_t) is explicitly conditioned on observations. While this could avoid approximate inference time guidance, it also makes the proposal dependent on the observation configuration. The generalization of the learned proposal under observation distribution shift remains unclear.

Questions for authors
  1. At what point or under what conditions should one transition from FPPF to L-FPPF?
  2. Could the authors provide a runtime or computational cost comparison with the main baselines?
  3. How robust is FPPF to changes in the observation configuration at test time?
3 comments

Thank you for the positive assessment and the clear questions. We answer them below, with new results to be added in the revised PDF (changes in blue).

W1 / Q2. Runtime comparison with the main baselines.

We measured per-step inference time and peak memory on SQG (64×64×264 \times 64 \times 2, one A100) for all main baselines, and ran a matched wall-clock comparison. Full tables are in our response to Reviewer TU32. In brief:

MethodNNTime (s/step)Peak GPU memory (GB)
BPF2003.870.05
LBPF2003.850.12
LETKF507.860.07
FA-APF20021.501.90
DAISI2036.211.80
L-FPPF2005.742.26

L-FPPF costs 1.5x LBPF at equal NN, is faster than LETKF with N=50N = 50, and is 3.7x and 6.3x faster than FA-APF and DAISI, which sample with inference-time guidance. At equal wall-clock time (LBPF N=300N = 300, LETKF N=33N = 33), L-FPPF has the lowest RMSE and CRPS in three of four SQG observation settings. Training the L-FPPF proposal on SQG takes about 20 minutes on one A100, compared with about 5 hours for the FA-APF prior (for DAISI we used the released checkpoint). On Lorenz-96 (dx=20d_x = 20), every neural method trains in about 30 minutes. LBPF and LETKF need no training.

Q1. When to move from FPPF to L-FPPF.

When to switch from global FPPF to L-FPPF depends on state dimension and properties. Even optimal proposal filters suffer from degeneracy for high state dimensionalities, as we demonstrated in the Lorenz-96 experiments. L-FPPF resolves this by taking advantage of local dependence, so states must exhibit some of this spatial structure for the method to excel. In practice we suggest running global FPPF and monitoring its pre-resampling ESS. If the ESS falls to a small fraction of NN and keeps decreasing as observations accumulate, the filter is close to collapse and L-FPPF is the better choice.

W2 / Q3. Robustness to changes in the observation configuration.

This is indeed a downside of learning the proposal, which we will state more clearly in the limitations. When observations are merely low quality or there is some distribution shift, the impact is limited. The importance weights always use the inference time computed likelihood p(ot∣xt)p(o_t \mid x_t), so a proposal trained under a different configuration lowers the ESS but does not invalidate the filter, provided it covers the posterior.

However, when the observation operator itself changes, the proposal must be retrained, although it need not be from scratch. As demonstrated in the SEVIR experiment, only an observation adapter (200,960 parameters) can be trained on top of a frozen 30.08M-parameter forecaster. One forecaster could therefore serve several observation types, each with its own lightweight adapter, an extension we leave to future work.

References: Snyder, C., Bengtsson, T., and Morzfeld, M. Performance bounds for particle filters using the optimal proposal. Monthly Weather Review, 143(11):4750-4761, 2015.

Dear Reviewer Nqrp,

Thank you again for your positive review. As the discussion period is coming to a close, we wanted to check whether our responses have addressed your questions. In brief:

  • Computational cost (W1, Q2). We report per-step time and peak memory for all main baselines on SQG. L-FPPF costs 1.5× LBPF at equal NN, is faster than LETKF with N=50N=50, and is 3.7× and 6.3× faster than FA-APF and DAISI. It trains in about 20 minutes, against about 5 hours for the FA-APF prior. At matched per-step cost, L-FPPF has the lowest RMSE and CRPS in three of the four SQG observation settings.
  • When to use L-FPPF (Q1). Global FPPF suits systems without spatial structure. When its pre-resampling ESS falls to a small fraction of NN and keeps decreasing, L-FPPF is the better choice, provided the system has the local dependence that localization exploits.
  • Observation configuration (W2, Q3). Because the importance weights use the likelihood at inference time, a mismatched proposal lowers efficiency but does not invalidate the filter. When the observation operator changes, only a lightweight adapter needs retraining (0.67% of the forecaster's parameters in our SEVIR experiment). We now state this trade-off, and the assumption of paired historical observations in the writing.

We would be happy to address any remaining questions, and if our responses have resolved your concerns, we would be grateful if you would consider revising your assessment.

Thank you for the detailed responses and additional experiments. The runtime comparisons have addressed my concern about computational cost, and the explanations regarding L-FPPF and observation configuration shifts are helpful. While the robustness to observation shifts remains partially unresolved, I appreciate the authors' efforts and will maintain my original rating.

Reviewer_TU32OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Flow Proposal Particle Filters (FPPF), which use conditional flow matching to learn an observation-conditioned proposal approximating the locally optimal proposal (p(xt∣xt−1,ot)p(x_t\mid x_{t-1},o_t)). The proposal density is evaluated through the continuous change-of-variables formula and used in the SIR importance-weight update. To mitigate weight degeneracy in high-dimensional systems, the authors introduce L-FPPF, a localized approximation that computes block-specific weights and performs resampling independently within each block. The methods are evaluated against classical and generative data assimilation methods on Lorenz-96, Lorenz-63, SQG, and SEVIR.

Critical strengths
  1. An observation-informed proposal that preserves the flexibility of particle filtering. The method accommodates nonlinear dynamics and observation operators without imposing a Gaussian filtering approximation. The experiments demonstrate improved performance over bootstrap particle filtering in several settings and robustness to nonlinear observations and multimodal forecasts. This supports the practical value of learning the proposal directly.

  2. Effective scaling through localization. The localized method substantially mitigates the degradation of global particle filters as dimension increases. In particular, a local proposal trained on 25-dimensional Lorenz-96 transfers to systems with up to 1,000 dimensions without retraining. The SQG and SEVIR experiments provide additional evidence that the approach can operate on high-dimensional spatial fields.

Critical weaknesses
  1. The computational tradeoff is insufficiently evaluated. FPPF requires iterative flow sampling and density evaluation, including derivative computations. Appendix C.4 reports 32 Euler steps for sampling and another 32 for likelihood evaluation. These operations suggest potentially substantial inference overhead, but no wall-clock comparison is provided. Consequently, it is difficult to determine whether the accuracy gains justify the additional cost relative to LBPF and LETKF.

  2. The gains over strong localized baselines are sometimes modest. Although the improvements over global particle filters are substantial, the advantage over LBPF is relatively small on SQG and SEVIR, and LETKF performs best on SQG. These results support the method’s usefulness, but leave uncertainty about the magnitude and breadth of its advantage over established localized methods, especially when computational cost is considered.

Questions for authors
  1. Computational efficiency. Could you report inference time and peak memory usage for FPPF/L-FPPF, LBPF, and LETKF on the same hardware, together with the particle or ensemble counts? A comparison of accuracy at matched wall-clock budgets would materially help assess whether the proposed method offers a favorable practical tradeoff. Please also report the training cost separately.

  2. Contribution and accuracy of importance correction. Using the same trained proposal, could you compare:

  • Samples used without importance correction;

  • Importance correction using deterministic exact-trace density evaluation, with numerical integration error controlled;

  • Importance correction using Hutchinson trace estimation, preferably with different probe counts?

This comparison would help establish how much of the improvement comes from the observation-conditioned proposal and how much comes from importance correction. It would also clarify the practical effect of stochastic density evaluation: an unbiased log-density estimate does not generally yield an unbiased importance weight after exponentiation. Small-dimensional experiments where exact-trace evaluation is feasible would be sufficient to address this question.

4 comments

Thank you for the careful review. We respond to each of the comments below.

W1 / Q1. Inference time, peak memory, matched wall-clock accuracy, training cost.

Table A gives the average cost per assimilation step on SQG (64×64×264 \times 64 \times 2), measured on a single A100 (40 GB), as well as the ensemble sizes (NN) used by each method in our SQG experiments.

MethodNNTime (s/step)Peak GPU memory (GB)
BPF2003.870.05
LBPF2003.850.12
LETKF507.860.07
FA-APF20021.501.90
DAISI2036.211.80
L-FPPF2005.742.26
  • BPF time is almost entirely the SQG forward model, which accounts for about two thirds of an L-FPPF step. Sampling the proposal and evaluating its density (32 + 32 Euler steps) adds 1.9 s, so L-FPPF costs 1.5x LBPF at equal NN.
  • LETKF with N=50N = 50 is slower than L-FPPF with N=200N = 200. Its cost is dominated by an N×NN \times N eigendecomposition at each of the 4096 grid points. We use the default batched PyTorch eigensolver, a specialized solver could be faster. However, LBPF and LETKF do not require a GPU.
  • L-FPPF is 3.7x faster than FA-APF and 6.3x faster than DAISI, which runs 10x fewer members. Both baselines sample with MMPS guidance, so every denoising step backpropagates through the full U-Net and solves a linear system with additional Jacobian-vector products. DAISI also runs 100 inverse and 100 forward steps. L-FPPF needs no guidance, and localization keeps its velocity network small.
  • Training costs: the L-FPPF proposal on SQG trains in about 20 minutes on one A100, while the FA-APF prior takes about 5 hours. This is because localization allows us to use a smaller network compared to the global FA-APF. For DAISI we used the released checkpoint, and its training cost is not reported. On Lorenz-96 (dx=20d_x = 20), every neural method trains in about 30 minutes on one A100. LBPF and LETKF need no training.

Matched wall-clock experiments. We thank the reviewer for this suggestion, which better helps to compare methods. We adjusted each baseline's ensemble size until its measured time per step matched L-FPPF with N=200N = 200 (5.74 s). This gives LBPF N=300N = 300 and LETKF N=33N = 33. DAISI would only fit 1 member in this budget, so we instead report it at its default N=20N = 20 (the ensemble size used in Andrae et al.), which is 6.3x the L-FPPF budget. The four SQG observation settings follow DAISI (Andrae et al., 2026) and were added at the request of Reviewer E2hN. Setup details are in our response to E2hN. Noisy is the original setting in the submission.

Table B. SQG, RMSE / CRPS. Bold marks the best matched-budget method. Standard deviations are in Appendix.

MethodNNNoisySparseSaturatingMultimodal
Matched budget (~ 5.74 s/step)
L-FPPF2003.259 / 1.7973.135 / 1.6541.758 / 0.9193.044 / 1.608
LBPF3003.235 / 1.7964.841 / 2.5732.016 / 1.0296.373 / 3.467
LETKF332.924 / 1.5924.962 / 2.6394.922 / 2.7744.020 / 2.199
Default configuration
LBPF2003.271 / 1.8124.890 / 2.6042.061 / 1.0536.556 / 3.579
LETKF502.86 / 1.533.694 / 1.9224.588 / 2.6363.023 / 1.625
DAISI204.509 / 2.6993.386 / 1.7232.804 / 1.4553.688 / 1.812
BPF20010.046 / 7.99210.189 / 8.18610.244 / 8.21510.116 / 8.098

W2. Gains over LBPF and LETKF.

We agree for the SQG setting in the submission (Noisy), this setting has a linear observation operator with Gaussian noise, which matches the assumptions of LETKF. The observations are also weakly informative due to the high noise, so the optimal proposal is closer to the transition density and the bootstrap proposal used by LBPF loses less.

The three added settings show the gain when the observation is more informative or nonlinear. Against LBPF at matched budget, L-FPPF lowers RMSE by 35% (Sparse), 13% (Saturating), and 52% (Multimodal). Across all four settings, L-FPPF is within 1.14x of the best RMSE, while each baseline is at least 1.6x from the best in some setting: LBPF 2.2x (Multimodal), LETKF 2.6x (Saturating), DAISI 1.6x (Saturating). This agrees with high-dimensional Lorenz-96, where L-FPPF lowers CRPS by about 25% relative to LBPF under arctan⁡\arctan and by a factor of 2.8 relative to LETKF under min⁡(x4,10)\min(x^4, 10) (Tables 3 and 4).

We agree that on SEVIR the margin over LBPF is smaller by comparison. However, the experiment importantly shows that our framework works when neither the dynamics nor the observation model is known, and supports a cross-modal observation setting that the guidance-based baselines do not support without substantial adaptation. Better proposal architectures and observation conditioning for real-world data will be an important component for improving this and are left for future work.

Q2. Contribution and accuracy of the importance correction.

Table C uses the same trained proposal in every column (Lorenz-96, h=arctan⁡h = \arctan, 100 trajectories). "No correction" uses the proposal samples with uniform weights. Entries are RMSE / CRPS.

dxd_xNNNo correctionExact traceHutchinson K=1K{=}1K=4K{=}4K=16K{=}16
510000.310 / 0.1680.187 / 0.1150.187 / 0.1150.188 / 0.1150.188 / 0.115
1010000.695 / 0.3770.288 / 0.1650.286 / 0.1640.284 / 0.1620.283 / 0.162

At dx=5d_x = 5 the importance weight correction lowers RMSE from 0.310 to 0.187; at dx=10d_x = 10 it lowers RMSE from 0.695 to 0.288. The uncorrected filter is also much less stable (RMSE standard deviation 0.36 vs 0.06 at dx=10d_x = 10).

Hutchinson vs exact trace. Every KK is within 0.005 RMSE of the exact trace, well below the standard deviation across trajectories. A single probe is also about 3x faster than the exact trace at dx=10d_x = 10. Appendix A.1 already measured the estimator itself, where its standard deviation in log⁡qϕ\log q_\phi is below 0.05 for dx≤25d_x \leq 25 and 0.247 at dx=50d_x = 50 with K=1K = 1.

Bias after exponentiation. We agree that an unbiased log-density estimate does not generally yield unbiased importance weights. Table C confirms the effect on accuracy is negligible where we use global FPPF. All L-FPPF results don't involve this stochastic estimate.

Dear Reviewer TU32,

Thank you again for your review. As the discussion period is coming to a close, we wanted to check whether our responses have addressed your concerns. In brief:

  • Computational cost (W1, Q1). We report per-step time and peak memory for all methods on SQG. L-FPPF costs 1.5× LBPF at equal NN, is faster than LETKF with N=50N=50, and is 3.7× and 6.3× faster than FA-APF and DAISI. It trains in about 30 minutes on one A100.
  • Matched wall-clock budget (Q1). At equal per-step cost, L-FPPF has the lowest RMSE and CRPS in three of the four SQG observation settings.
  • Gains over localized baselines (W2). In the added SQG settings, L-FPPF lowers RMSE relative to LBPF by 35% (Sparse), 13% (Saturating) and 52% (Multimodal) at matched cost.
  • Importance correction (Q2). Using the same proposal, the correction lowers RMSE by 40% at dx=5d_x=5 and 59% at dx=10d_x=10. Hutchinson estimation with a single probe is within 0.005 RMSE of the exact trace, and L-FPPF removes the need for this.

All results are also included in the revised PDF. We would be happy to address any remaining questions, and if our responses have resolved your concerns, we would be grateful if you would consider updating your assessment.

Thank you for your thorough and constructive responses to my review comments.

I appreciate the additional computational cost analysis, which demonstrates that the proposed L-FPPF is substantially faster than FA-APF and DAISI, benefiting from localization and the resulting reduction in computational complexity.

I also find the matched wall-clock comparisons particularly valuable. The additional experiments provide useful insights into how the relative performance of different filtering methods depends on the observation settings. However, since these experiments are conducted primarily on the SQG system, some caution is needed when generalizing the findings to broader classes of state-space models. I believe the results support positioning L-FPPF not as a universally superior method, but as a particularly promising approach for challenging observation settings, especially those involving nonlinear observation operators or conditions where conventional Gaussian approximations may be inadequate.

Furthermore, the additional ablation study provides convincing evidence that importance weight correction substantially contributes to the filtering accuracy. I consider this an important result that strengthens the methodological contribution of the paper.

Overall, I believe the authors have adequately addressed my major concerns, and the additional experiments have significantly strengthened the paper. I appreciate the authors' efforts and have therefore decided to raise my rating to 3: Weak Acceptance.

Comments on the paper

1 comment

Summary of Revisions and Responses to Reviewers

We sincerely thank all reviewers for their constructive feedback and discussion. During the rebuttal, we substantially strengthened the paper through experiments, additional baselines, and clarifications. Below, we summarize how we addressed the main concerns raised during the review. Changes made during this period are highlighted in blue in the revised manuscript.

  • Novelty and fair comparisons (E2hN, CZb1): We clarified our contributions relative to existing flow based particle filters and revised any claims to reflect this. To address concerns about training with observations, we added comparisons with NF-DPF and an observation-conditioned RealNVP trained on the same data, along with a Kalman-informed proposal baseline. FPPF consistently outperforms these alternatives, providing stronger evidence for the benefit of our proposal-learning approach.

  • Improvements beyond localization (E2hN, CZb1, TU32): The original SQG experiment showed limited improvement over LBPF. We therefore expanded the evaluation to more challenging observation settings, where both methods use the same localization but L-FPPF reduces RMSE by 13–52%. These results demonstrate that the learned proposal provides substantial benefits beyond localization, particularly when observations are informative or nonlinear.

  • Computational cost versus accuracy (TU32, Nqrp, CZb1): We added detailed runtime, memory, and training-cost comparisons, including experiments at matched inference budgets. While L-FPPF is more expensive than LBPF per particle, it achieves the best accuracy in three of four SQG settings at matched wall-clock cost. This provides a clearer picture of when the additional computation is justified.

  • Importance correction and density estimation (TU32): We isolated the contribution of importance weighting by comparing the same trained proposal with and without correction. Removing the correction substantially degrades filtering accuracy, while stochastic Hutchinson estimates perform nearly close to exact trace evaluation for low dimensional settings. This establishes the importance and effectiveness of the Bayesian correction step.

  • Robustness and adaptation (CZb1, Nqrp, E2hN): We investigated whether errors in previous-state particles undermine a proposal trained on true trajectories. Controlled perturbation experiments show that FPPF remains effective under such errors, consistent with its stable long-term cycling performance. We also clarified the dependence on observation configuration and clarified the adaptation of a frozen generative forecaster through a lightweight observation adapter on SEVIR.

  • Reproducibility and limitations (E2hN): We verified the DAISI implementation, closely reproduced its results in three of four SQG settings, and explicitly documented the remaining discrepancy. We also clarified ensemble-size comparisons, added spatial visualizations and spectral diagnostics, and discussed the spatial-coherence limitations shared by localized particle filters.

We believe these revisions address the central questions about novelty, practical benefits, computational tradeoffs, and robustness, while providing a clearer account of the method's assumptions and limitations.

We also appreciate that three of the four reviewers engaged in follow-up discussion, and that all three have now expressed support for acceptance.

We thank the reviewers and Area Chair again for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.