- 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 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
W2 / Q3-Q4. Expanded SQG experiments and DAISI reproduction.
As requested, we expanded SQG to all four observation settings of DAISI [4]: Noisy (, , 25% observed), Sparse (, 1, 5%), Saturating (, 0.01, 25%) and Multimodal (, 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):
| DAISI | Noisy | Sparse | Saturating | Multimodal |
|---|---|---|---|---|
| Reported in [4] | 2.46 / 1.32 | 3.40 / 1.73 | 2.96 / 1.54 | 3.71 / 1.81 |
| Ours | 4.509 / 2.699 | 3.386 / 1.723 | 2.804 / 1.455 | 3.688 / 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):
| Setting | L-FPPF () | LBPF () | DAISI () | LETKF () |
|---|---|---|---|---|
| Noisy | 3.259 / 1.797 | 3.271 / 1.812 | 4.509 / 2.699 | 2.86 / 1.53 |
| Sparse | 3.135 / 1.654 | 4.890 / 2.604 | 3.386 / 1.723 | 3.694 / 1.922 |
| Saturating | 1.758 / 0.919 | 2.061 / 1.053 | 2.804 / 1.455 | 4.588 / 2.636 |
| Multimodal | 3.044 / 1.608 | 6.556 / 3.579 | 3.688 / 1.812 | 3.023 / 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 , but L-FPPF has 27% to 30% lower CRPS at every (0.156 to 0.162 vs 0.222 to 0.224). Under , L-FPPF lowers RMSE by 41% to 47% and CRPS by about half relative to DAISI at every 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 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 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 -adjusted fair CRPS to avoid biased comparisons.
- I'm also curious how well LFPPF would work with a smaller 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.
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.
NF-DPF improves on BPF mainly at dx≥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=50 (RMSE 0.323 vs 0.659 under arctan, 0.519 vs 1.687 under min(x4,10)). 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.