acceptodds
Under review as a conference paper at ICLR 2027

Bayesian Filtering in Physical Systems via Test-time Trained Flow Matching

Abstract

Bayesian filtering provides a principled framework for online state estimation under uncertainty. Applying it to high-dimensional systems with complex posterior distributions remains challenging. Recent generative models, such as flow matching, have shown potential in Bayesian filtering. However, existing approaches either propagate a finite sample ensemble or repeatedly sample a state trajectory, coupling online cost to the ensemble size or trajectory sampling budget. To address this, we propose a new perspective of directly encoding the evolving distribution into flow matching model weights, namely, the Belief Flow Filter (BFF). It is a generative filtering framework that updates model weights via gradient descent at test time to track the posterior evolution. BFF carries the belief forward in generative model weights and draws samples from the updated model without restricting the posterior to a Gaussian family. A Bayes-derived loss motivates the update design, where an idealized expressivity result and a conditional, population-level bound connect the outer training loss to filtering error. The deployed finite-capacity update amortizes a specified observation model and has no guarantee of exact Bayesian inference under an unseen likelihood. Across five physical systems, BFF achieves the lowest reconstruction error and continuous ranked probability score (CRPS) on Kuramoto–Sivashinsky (KS) and Burgers' and ranks among the top two methods on all four metrics for Navier–Stokes (NS) flow with uniform sensors. BFF also achieves the lowest reconstruction error, spectrum error, and miscalibration area (MA) for NS flow with a single moving sensor, and the lowest gradient error, spectrum error, and MA for synthetic tokamak profile estimation from diagnostic measurements.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Bayesian Filtering, Flow Matching, Test-time Training, State Estimation

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 3 reviews, above 85% of the others.
Reviewer_2RYrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes the Belief Flow Filter (BFF) for high‑dimensional physical‑system state estimation via Bayesian filtering. It propagates filtering belief through the weights of a persistent generative model and draws state samples from the updated generator. Experimental results demonstrate competitive reconstruction, spectral accuracy, and uncertainty calibration across physical systems with chaotic dynamics, sparse observations, and nonlinear observation models.

Critical strengths
  1. It is interesting to store the sequential filtering belief within generator weights. It offers an alternative route to recursive Bayesian filtering for high‑dimensional chaotic PDE systems, mitigating the curse‑of‑dimensionality suffered by particle‑based methods.
  2. The paper evaluates across multiple chaotic physical tasks covering 1D-PDE, 2D-PDE, low-dimensional ODE and a realistic-inspired tokamak simulation.
Critical weaknesses
  1. A primary motivation for BFF is reducing online cost by amortizing filtering updates in offline training. While the TTT weight‑update step is fast, inference runtime is still dominated by flow‑ODE‑based posterior sampling. A more systematic cross‑method computational comparison is desired, including training overhead, per‑observation update time, sampling latency and memory consumption. It should also be clarified whether reported timings adopt matched hardware, batch size, accuracy requirements and sampling budgets.
Questions for authors
  1. The proposed method relies on a learned surrogate loss for online test-time updates. How does the surrogate loss approximate the ideal Bayesian filtering objective, and can the authors provide quantitative evidence that minimizing this surrogate leads to improved posterior estimates?

  2. The theoretical analysis appears to apply to general recursive belief encoders, rather than specifically establishing the advantages of the proposed test-time training mechanism. What theoretical guarantees distinguish BFF from conventional recurrent belief encoders, such as GRUs?

  3. The method requires test-time parameter updates and flow-based posterior sampling. Could the authors provide a comprehensive comparison of training cost, online update latency, posterior sampling time, and memory consumption against conventional Bayesian filtering methods and other learned filtering approaches?

  4. The evaluation primarily uses metrics such as relative error and CRPS. How do the authors verify that BFF captures the joint posterior distribution, particularly in high-dimensional systems, rather than merely producing accurate point estimates or marginal distributions?

9 comments
AuthorsOpenReview ↗

Official Response to Reviewer 2RYr (1/3)

We thank the reviewer for the detailed and constructive review comments. We are encouraged that the reviewer recognizes that our method is interesting, is a promising alternative to conventional particle-based filters, and is evaluated in different physical tasks. Next, we address the concerns of the reviewer one by one:

Comment1 (W, Q3): A comprehensive computational comparison including training overhead, per‑observation update time, sampling latency and memory consumption is desired. It should also be clarified whether reported timings adopt matched hardware, batch size, accuracy requirements, and sampling budgets

We thank the reviewer for this comment. We conducted a new benchmark of all methods reported in Table 1 NS and KS. Below, we summarize the benchmark results on the NS experiment:

MethodSampling stepsTotal Training GPU HoursPer-observation update (ms)sampling latency (ms)Update + Sampling (ms)Peak memory usage (GB)
ENKF-015360.11536.10.063
EnSF-01996.3150.52146.80.068
FlowDAS1003.4--1908.33.152
SDA10022.7--2190.44.061
FNO-9.913.2013.21.334
Transformer-7.17.807.80.619
BFF10033.512.9830.7843.60.459
BFF133.512.99.422.30.459

The time is measured on a single H800 GPU. We did not force the methods to match accuracy or sampling budget requirements. Instead, we use configurations that produce Table 1 results so that the benchmark corresponds to the actual experiments. The inference batch size is 1 except for EnKF and EnSF because they need ensemble updates inherently. For SDA and FlowDAS, their update and sampling steps are coupled, so we only report a combined time.

These results demonstrate that the update of BFF is fast, comparable to deterministic methods like FNO or Transformer. BFF's inference time can be dominated by the ODE integration. Therefore, it can be accelerated if its performance does not drop too much when the sampling steps decrease. We show that this is possible in NS (Table 4 in the revised manuscript) and the Tokamak experiment (Table 6 in the revised manuscript).

We have added the complete benchmark results and detailed configurations into Appendix J.1 in the revised manuscript.

Comment2 (Q1): The proposed method relies on a learned surrogate loss for online test-time updates. How does the surrogate loss approximate the ideal Bayesian filtering objective, and can the authors provide quantitative evidence that minimizing this surrogate leads to improved posterior estimates?

We thank the reviewer for this question. We agree that directly assessing the correlation between the surrogate loss and the ideal Bayesian filtering loss would strengthen the validity of our method design.

First, we clarify that we did not make rigorous theoretical claims in the manuscript that the learned surrogate loss converges to the ideal Bayesian filtering loss. However, we did observe evidence suggesting that our learned loss is positively correlated with the Bayesian ideal filtering loss, and the directions of their gradients are aligned in the high-dimensional weight space.

Because the ideal loss in Eq. (9) is intractable in high-dimensional space as discussed in Section 2.3, we did additional diagnostic experiments on the 3-dimensional Lorenz 63 system. There, we can approximate the ideal filtering loss by sampling from the model posterior, propagating, reweighting with the likelihood, and then approximating the ideal marginal velocity field uKbθtu_{Kb_{\theta_t}} using Monte-Carlo estimation (please see Eq. 77, Appendix J.11 of the revised manuscript, and also, e.g., Eq. 12 in [1]).

With this approximated uKbθtu_{Kb_{\theta_t}}, we can approximate the ideal loss ℓ⋆\ell^\star (Appendix J.11 Eq. 79), and do diagnostics. Please see the details in Appendix J.11.1 in the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer 2RYr (2/3)

(Response to Comment 2, continued)

We first compare, across 20 trajectories, the time series of the surrogate loss and the ideal filtering loss, and observe significant correlation between them.

MetricPooled PearsonPooled Spearman
Estimate0.64820.3707
95% confidence interval[0.6021, 0.6893][0.3344, 0.4050]
Pairing-test P2×10−62\times 10^{-6}2×10−62\times 10^{-6}

They have a pooled Pearson correlation of ~0.65, indicating moderate to strong correlation, and both correlations have a positive 95% CI. Both p-values reach the Monte Carlo resolution, showing that the surrogate loss and the ideal filtering loss have a statistically significant correlation in this experiment.

We then examine whether the learned negative gradient is a local descent direction for the ideal filtering loss. The gradient inner product is positive at 1,974 of 1,980 update contexts (99.70% of the steps, 95% CI: [99.29%, 100%]). Therefore, a sufficiently small gradient descent step on the learned surrogate loss would decrease the ideal filtering loss, which can be associated with the distributional posterior error using a Grönwall argument (Appendix F.4 in the revised manuscript). We have added these to Appendix J.11 in the revised manuscript. Please see that section for details of the experiment settings and results.

We also give a possible intuitive explanation to how the surrogate loss approximates the ideal Bayesian filtering objective. Given the update rule in Eq. (12), θt+1=θt−η∇θℓϕ(θ;θt,ot+1,at)\theta_{t+1} = \theta_{t} - \eta \nabla_\theta \ell_\phi(\theta;\theta_t,o_{t+1},a_t). Meanwhile, the ideal update is θt+1⋆=arg⁡min⁡θℓ⋆(θ)\theta_{t+1}^{\star} = \arg \min_{\theta} \ell^\star(\theta) where ℓ⋆\ell^{\star} is the ideal filtering loss.

Suppose we train the surrogate parameter ϕ\phi to make the model's updated weight θt+1\theta_{t+1} approach the oracle updated weight θt+1⋆\theta_{t+1}^\star, the negative gradient of the learned surrogate loss −η∇θℓϕ- \eta \nabla_\theta \ell_\phi would approach the displacement toward the oracle minimizer arg⁡min⁡θℓ⋆−θt\arg \min_{\theta} \ell^\star - \theta_t. Then, if the ideal loss is well-shaped (e.g., convex) and near θ\theta so that its negative gradient is positively correlated with the displacement toward the minimizer, the gradients of the learned loss and the ideal loss can therefore be positively aligned.

We leave a more rigorous theoretical treatment for future work.

[1] Lin et al., On the Design of One-step Diffusion via Shortcutting Flow Paths, ICLR'26.

Comment3 (Q2): The theoretical analysis appears to apply to general recursive belief encoders, rather than specifically establishing the advantages of the proposed test-time training mechanism. What theoretical guarantees distinguish BFF from conventional recurrent belief encoders, such as GRUs?

We thank the reviewer for this comment. The theoretical result in Section 3.2 shows that our outer training loss controls the distributional error in Bayesian filtering. We agree that it also applies to conditional flow matching/diffusion filters that are conditioned on a latent state updated with recurrent models. We view a BFF-specific theoretical guarantee on convergence or a larger model capacity as an interesting future direction.

Meanwhile, we would like to clarify that theoretical results in the original manuscript consist of two major parts. In addition to the results mentioned above (Section 3.2 in the revised manuscript), we also derived the weight space update rule for Bayesian filtering (Section 2.3 and 3.1 in the revised manuscript). This theoretical result guides us to design the weight-space gradient update and the specific form of the surrogate loss in our BFF method, instead of using a generic recurrent belief encoder.

We also provide empirical evidence in our new experiments to show that this design is better than GRU or a generic gradient update. Please see our results in Appendix J.2 and J.10 in the revised manuscript, or our response to Reviewer STer comment 1 and VhJ9 comment 2.

AuthorsOpenReview ↗

Official Response to Reviewer 2RYr (3/3)

Comment4 (Q4): The evaluation primarily uses metrics such as relative error and CRPS. How do the authors verify that BFF captures the joint posterior distribution, particularly in high-dimensional systems, rather than merely producing accurate point estimates or marginal distributions?

We thank the reviewer for this insightful question. We agree that evaluating the distributional accuracy of the joint posterior distribution is important.

We clarify that the metrics in the original manuscript span sample-wise accuracy (field relative L2 and spectrum deviation) and distributional calibration and sharpness (MA and CRPS). A point estimate would fail on distributional metrics, while a distribution marginalized over history-information conditions would be bad on sample-wise accuracy. However, these metrics are indeed less sensitive to spatial dependence of the posterior.

Therefore, to further test whether BFF learns the joint posterior or a spatially independent marginal distribution, we designed a spatial-dependence diagnostic on the NS uniform and NS moving-sensor (both are 128×128128 \times 128 dimensional) benchmarks.

We use the p=1 variogram score (VS), which measures how well the model posterior predicts actual samples' spatial differences, and thus reflects spatial dependence on the posterior distribution:

VSij=(∣yi−yj∣−1M∑m=1M∣xi(m)−xj(m)∣)2,VS_{ij} = (|y_i-y_j| - \frac{1}{M}\sum_{m=1}^{M} |x_i^{(m)}-x_j^{(m)}| )^2,

where yiy_i and yjy_j are actual field values at locations ii and jj, and xi(m)x_i^{(m)} and xj(m)x_j^{(m)} are the field values of the mm-th sample in the model posterior ensemble conditioned on the same history observations.

To test that BFF captures the joint spatial distribution, we compare the VS of the spatially independent marginal distribution with that of the original joint posterior. We approximate the model's spatially independent marginal (i.e., product of pointwise marginal distributions) by randomly shuffling the sample index for each selected location. If the original joint posterior distribution has a VS lower than that of the marginal distribution, then we have evidence that BFF learns a meaningful joint spatial distribution.

We summarize the results in the table below (4 displayed decimals). Please see Appendix J.6.1 for more detailed configuration of the evaluation.

NS moving sensorNS moving sensorNS moving sensorNS moving sensorNS uniformNS uniformNS uniformNS uniform
Spatial distanceOriginal VS ↓Shuffled VS ↓Rel. change after shuffle95% CI for relative changeOriginal VS ↓Shuffled VS ↓Rel. change after shuffle95% CI for relative change
1 pixel0.01420.18991233.99%[816.46%, 1761.40%]0.01310.0515+294.73%[160.98%, 439.62%]
4 pixels0.08800.162484.52%[53.40%, 123.36%]0.04670.0571+22.17%[11.98%, 30.03%]
16 pixels0.40020.40631.53%[0.006%, 3.19%]0.14370.1436−0.06%[−1.50%, 1.10%]
32 pixels0.55390.55470.14%[−0.68%, 0.86%]0.15680.1567−0.10%[−0.73%, 0.27%]

We observe that for the moving sensor or uniform sensor NS scenarios, the spatial joint posterior has a lower VS than the marginal one at spatial distances ≤16\le 16 pixels for moving sensors and ≤4\le 4 pixels for uniform sensors. We note that for uniform sensors, the sensor is placed with a spacing of roughly 128/8=16128 / 8 = 16 pixels, so the spatial correlation of the posterior uncertainty conditioned on the sensor observation would decay in this interval. In contrast, there is only one sensor in the moving sensor setting, which allows larger spatial regions for uncertainty.

Additionally, we compared the VS of BFF with the posterior modeling baselines in the NS uniform setting, averaged over the four pair distances (1, 4, 16, 32).

MethodOriginal VS ↓Shuffled VS ↓Relative change after shuffle
BFF0.090090.10224+13.49%
SDA0.053320.0560+5.04%
FlowDAS1.57692.7891+76.87%
EnSF0.58090.5827+0.31%
EnKF0.43040.4304−0.00%

Compared to the baselines, the original VS of BFF is the second best. We note that SDA is underconfident, with a high miscalibration area ~0.43 on NS (see Table 1 in the manuscript), and it makes predictions with less spatially correlated fluctuations. This is reflected in its smaller relative VS change after shuffling. Meanwhile, FlowDAS has a worse VS absolute value. EnSF and EnKF demonstrate little relative VS increase after shuffling, suggesting their posteriors are not much more informative than the corresponding spatially independent marginals.

These results indicate that BFF's posterior captures a reasonable spatially joint distribution. We have added these results in Appendix J.6.1 in the revised manuscript.

Reviewer_VhJ9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Belief Flow Filter (BFF), a Bayesian filtering framework that represents the evolving posterior distribution through the weights of a flow matching model. Instead of maintaining particle ensembles, BFF recursively updates the generator's weights using a learned surrogate loss. The update model is meta-trained through an unrolled conditional flow-matching loss. Experiments across several physical systems demonstrate competitive reconstruction accuracy and uncertainty calibration, including promising generalization beyond the training horizon.

Critical strengths
  • interesting and new pespective on Bayesian filtering: this work proposes to represent the evolving distribution by the evolving weights of a flow matching model. This is an interesting alternative to particle-based filtering approch.
  • promising empirical results: the method performs competitively across several nonlinear physical systems, including settings with sparse observations and multimodal posteriors.
  • long-horizon evaluation beyond training horizon: The Lorenz-63 experiments demonstrate that the learned recursive update can remain effective significantly beyond the training horizon.
Critical weaknesses
  • A simple baseline to represent the Bayesian posterior distribution bt(s)b_t(s) is to train a single conditional flow matching model p(st∣ht)p(s_t|h_t) given all histories. The experiment does not seem to show any comparison to this alternative. It would strengthen the paper if it can show the weight-update method is better than a single conditional model in terms of performance-efficiency tradeoffs, given both of them require ODE integration from Gaussian noises.
Questions for authors
  • In the paper, the weight update policy is parameterized via the gradient of a specifically-designed loss term (equations 11,12). Are there any insights for parameterizing the policy in this way? For instance, how is it compared to simply parameterzing the update policy as a fully connected neural net (parameter in, parameter out), or as a gradient of a generic scalar-output neural net?
2 comments
AuthorsOpenReview ↗

Official Response to Reviewer VhJ9 (1/2)

We thank the reviewer for the constructive feedback. We are glad that the reviewer recognizes that our proposed method is an interesting alternative to particle-based Bayesian filters. We also appreciate that the reviewer finds empirical evidence competitive and promising. We address each concern of the reviewer below:

Comment1 (W): can the authors compare a flow matching given all history information in terms of performance-efficiency tradeoff?

We thank the reviewer for raising this question. We agree that comparing a CFM with history conditions is an important ablation study.

We clarify that the existing GRU ablation in Appendix J.2 of the revised manuscript already examines a history-conditioned flow model. However, that history is explicitly compressed into the the hidden states maintained by the recurrent module. Therefore, we conducted additional experiments that use a Transformer (with KV caching) to directly encode full history observation and action sequences.

We evaluate this alternative in the two NS settings with moving sensor or uniform sensors, with matched inference sampling compute. The inference time benchmark is performed on the same machine with a H800 GPU.

NS moving sensorNS moving sensorNS uniform sensorNS uniform sensor
QuantityBFFHistory Transformer-CFMBFFHistory Transformer-CFM
Relative L2 ↓0.37710.733960.22170.41537
Spectrum deviation ↓0.02860.067230.02860.03627
MA ↓0.17990.173770.11430.20967
CRPS ↓0.190830.334570.07660.13937
Observation update (s) ↓0.0360.0020.0090.001
Posterior sampling, 16 samples (s) ↓5.7265.4581.6461.640
Update + sampling (s) ↓5.7615.4601.6551.641
Peak allocated sampling memory (MiB) ↓1246.051063.69486.18414.58

In the moving-sensor experiment, BFF achieves lower relative L2, spectrum deviation, and CRPS, while in the uniform sensor case, BFF performs the best on all 4 metrics here. We attribute the performance gain of BFF to the explicit recursive Bayesian filtering structure. As the reviewer points out, since both methods require ODE integration during sampling, which dominates wall time. Thus, although the update step of BFF is longer, the total inference time (update + sampling) difference remains lower than 6%.

In summary, BFF demonstrates better performance-efficiency trade-off than an all-history-conditioned flow matching model in this experiment. We have added this experiment to Appendix J.8 of the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer VhJ9 (2/2)

Comment2 (Q): why is does the surrogate loss take the specific form? Does a weight space update using a parameter in-parameter out neural network or a gradient of a generic scalar function?

We thank the reviewer for the insightful question. We agree that this is also a meaningful ablation study.

First, we would like to clarify that the surrogate loss design was motivated by the theoretical analysis in Section 2 and Section 3.1 of the revised manuscript. We derived the ideal form of weight space Bayesian filtering updates (minimizing the ideal loss in Equation 9), which naturally motivates us to parameterize the update with the specific structure (gradient descent on surrogate loss E∥fϕ(θ)−gϕ(θ,θˉ,o,a)∥2\mathbb{E}\|f_\phi(\theta) - g_\phi(\theta,\bar{\theta},o,a)\|^2 in Equation 11).

Second, we did additional experiments to compare BFF with the two alternatives that the reviewer mentioned, i.e., parameter update with a fully-connected neural network θ←fϕ(θ)\theta \leftarrow f_\phi(\theta), or with a gradient descent on a parameterized scalar loss function θ−θ←∇θfϕ(θ)\theta - \theta \leftarrow \nabla_\theta f_\phi(\theta). The results in the NS uniform sensor setting are summarized as follows:

Update policyTotal parameters (M)Relative L2 ↓Spectrum deviation ↓MA ↓CRPS ↓
BFF (gradient update on structured surrogate loss)70.980.22170.02860.11430.0766
Fully connected MLP weight update152.380.94320.15100.24050.4893
Gradient update on generic scalar-network70.990.41010.04020.15800.1490

We observe that BFF still delivers the best performance among these variants on all four reported metrics. We note that fully connected MLP needs a large number of weights (~115% more than BFF) to directly process the weights. Therefore, they have a large parameter count and less competitive performance. Gradient updates on generic scalar network shows more competitive performance. BFF still outperforms it, which we attribute to the inductive bias of the Bayesian filtering problem in BFF.

These comparisons favor BFF's structured surrogate-gradient parameterization among the tested alternatives. We have added this experiment into the revised manuscript, the appendix J.10.

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

The paper introduces the Belief Flow Filter (BFF), a recursive generative filtering framework that carries the current belief in a subset of flow-model weights. A learned surrogate loss updates these weights with one test-time gradient step per observation, while an unrolled conditional flow-matching objective trains the update from state trajectories. The authors motivate the surrogate through a Bayes-derived propagation–likelihood decomposition and provide idealized expressivity and population-level error results. They evaluate BFF on Lorenz-63, Burgers’, Kuramoto–Sivashinsky, Navier–Stokes, and synthetic tokamak systems, covering sparse and moving sensors, nonlinear observations, calibration, CRPS, training-seed sensitivity, and observation-protocol shifts.

Critical strengths
  1. A substantive construction. Carrying belief in online-updated flow-model weights, with a Bayes-derived surrogate for the test-time update, is a genuine conceptual and technical contribution rather than a simple architecture tweak. It directly addresses the cost of repeatedly solving a trajectory-level inverse problem.This is consequential because it makes the core idea potentially publishable if the attribution and protocol issues below are resolved.

  2. A broad and candid evaluation. The paper covers five systems, a common endpoint protocol, CRPS and EnSF, a five-seed KS diagnostic, GRU/probe studies, observation shifts, and negative results. This is consequential because it gives the empirical claims credibility and makes the paper useful to readers even where the mechanism attribution is still open.

Critical weaknesses
  1. The central weight-space advantage is not yet isolated. For one online matrix W∈R3H×HW\in\mathbb{R}^{3H\times H} with a fixed probe xx, the inner loss depends on WW only through WxWx, so the update is rank-one per probe:

∇Wℓ=vx⊤,Wt−W0=ztx⊤,rank⁡(Wt−W0)≤1.\nabla_W\ell=v x^\top,\qquad W_t-W_0=z_t x^\top,\qquad \operatorname{rank}(W_t-W_0)\le 1.

With the default single probe, the online variation has at most 3HL=4,6083HL=4{,}608 dynamic coordinates, versus 3H2L=1,179,6483H^2L=1{,}179{,}648 nominal weights; with kk probes, the update has rank at most kk. The GRU control shares the frozen DiT backbone and training data, but the paper does not report the GRU hidden-state dimension or match this carried-state dimension; the probe sweep’s participation rank is only 1.86 at k=8k=8. In addition, the high-dimensional evaluation is marginal and does not directly test joint posterior structure. This is consequential because the current comparison cannot distinguish an advantage of weight-space memory itself from an advantage of a particular low-rank online adaptation; it therefore bears on whether the paper’s central scientific claim is established.

  1. The NS evaluation and main-table protocol are not yet clean. Appendix H.2 calls the use of test-split statistics a transductive preprocessing limitation, and H.5 confirms that NS scoring retains the test-split reference. Table 4 states 100 BFF samples for KS and 16 for NS, whereas H.5 and Table 7 indicate 16 for the KS models, and Table 7’s seed-0 metrics match Table 1’s BFF KS row at the reported precision. Appendix J.1 also reports different 100-to-2-step latencies and relative L2L_2 values from Table 4. This is consequential because NS is a central benchmark and the main table supports the headline claims; without a train-only normalization sensitivity and reconciled sample counts, the headline empirical comparison is not yet clean.
Questions for authors
  1. [Clarification/experiment] When rerunning inference, do the main NS results and method rankings remain stable if normalization uses only training-split statistics, while physical observations and noise realizations are held fixed?

  2. [Clarification/experiment] What is the GRU hidden-state dimension, and is the reported TTT advantage preserved when the carried recurrent-state dimension is matched or swept, with total parameter count reported separately?

  3. [Suggestion/future work] Can the authors provide one fixed-history or spatially joint diagnostic, beyond scalar-coordinate marginal CRPS and coverage, to show whether the high-dimensional posterior’s dependence structure is reasonable?

  4. [Clarification/sensitivity] What is the actual KS posterior sample count used in Tables 1, 4, and 7? With that clarified, which key cross-method differences remain stable under paired trajectory resampling and common-MM evaluation, with fair CRPS added only where its sampling assumptions apply?

6 comments
AuthorsOpenReview ↗

Official Response to Reviewer STer (1/4)

We thank the reviewer for the thoughtful and constructive feedback. We are encouraged that the reviewer finds that our method makes a genuine conceptual and technical contribution and that our evaluation is broad and candid. In the following, we address the reviewer's concerns one by one.

Comment 1 (W1 & Q2): BFF has a large carried state dimensionality in the weight space updates, but when compared to GRU hidden state updates, the hidden state dimension is not reported and may not be a fair comparison.

We thank the reviewer for raising this question. We agree with the importance of isolating the benefit of weight-space updates. Comparing BFF with GRU under controlled carried state dimensions and parameter count is thus important, and the original manuscript controlled the parameter count of GRU (a GRU-based CFM configured to have at least as many parameters as BFF).

Therefore, we did new experiments to test the performance of GRU with different hidden dimensions in the 2D NS uniform sensor setting, which can demonstrate whether the advantage of BFF persists when the carried-state dimension of GRU matches that of BFF.

Regarding the setting, we note that the actual BFF dynamic coordinates in the 2D NS uniform experiment are 3HL=18432 with H=512, L=12 (4608 in the reviewer's calculation is based on the default values reported in Appendix D). We adopted k=1 in BFF there because, as shown in the original Table 8, more probes tend to have a lower utilization.

Therefore, we swept the capacity of GRU carried states with the dimensions of 32, 128 (the originally reported GRU), 1024, 4096, 18432 (matches BFF update state dimension). The results are summarized in the table below:

MethodCarried-state dimensionTotal parametersRelative L2 ↓Spectrum deviation ↓MA ↓CRPS ↓
BFF18,43271M0.22170.02860.11430.0766
GRU3277M0.57320.06510.21710.2537
GRU12877M0.40290.03880.20480.1455
GRU1,02480M0.28420.03110.17830.0941
GRU4,096117M0.24730.03050.17700.0850
GRU, hidden dimension matched to BFF18,432852M0.56910.04710.20830.2309

We observe increasing the carried state dimension of GRU tends to improve its performance, which is expected because a larger capacity allows the posterior distribution to be stored more accurately. However, if GRU's dimension is increased to 18432 that matches BFF, its performance drops sharply. The best GRU uses about 4096 carried-state dimensions with about 65% more parameters than BFF, but BFF still outperforms it on all 4 metrics.

In standard GRUs, the weight matrix grows quadratically with the hidden dimension size. A high-capacity 18,000 dimensional hidden state that matches BFF would lead to a large total parameter count (over 800M).

Therefore, we further optimize GRU by dividing their hidden state into 4 or 16 groups for state dimensions 4096 or 18432, respectively:

MethodCarried-state dimensionTotal parametersRelative L2 ↓Spectrum deviation ↓MA ↓CRPS ↓
BFF18,43271M0.22170.02860.11430.0766
Grouped GRU4,09689M0.24790.02890.17510.0853
Grouped GRU, hidden dimension matched to BFF18,432135M0.23940.02870.17850.0836

This lowers the quadratic computation in GRUs and improves the performance of the GRU with matched hidden dimensions. Nevertheless, BFF still delivers the best performance despite having the lowest parameter count.

In summary, these results indicate that BFF is effective at utilizing the high-capacity state space through our weight-space TTT design. We added these results to Appendix J.2 of the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer STer (2/4)

Comment 2 (W2 & Q1): Is the result sensitive to normalization statistics?

We thank the reviewer for raising this question. We agree that using training-split normalization to run the evaluation is suitable. In the original Table 1, all methods were evaluated using test-split normalization statistics. We have now re-evaluated all checkpoints under training-split normalization and report the results below:

MethodEvaluation normalization statisticsRelative L2 ↓Spectrum deviation ↓MA ↓CRPS ↓
EnKFTraining split0.53520.02320.93700.3567
EnSFTraining split0.63990.16030.32040.4235
FlowDASTraining split1.27560.39590.23840.6751
SDATraining split0.15480.14570.43400.0688
FNOTraining split0.32230.23851.00000.1155
TransformerTraining split0.31400.12691.00000.1106
BFFTraining split0.22170.02860.11430.0766
BFFTest split0.22500.02860.13630.0747

The best and the second-best methods are highlighted in bold or italics, respectively. The original test-split normalization results of BFF are also included for reference.

We observe that most metrics are slightly improved, as model training uses training set statistics for normalization, and the test normalization is now aligned to training. The relative rankings between BFF and others remain unchanged, where BFF is the best or the second best across all four metrics. We have updated Table 1 in the revised manuscript accordingly.

Comment 3 (W1 & Q3): Can the authors demonstrate the spatial dependence of the high-dimensional posterior?

We agree that assessing the spatial dependence of the posterior is important. We therefore designed a spatial-dependence diagnostic on the NS uniform and NS moving-sensor benchmarks.

We use the p=1 variogram score (VS), which measures how well the model posterior predicts observed spatial differences, and thus reflects spatial dependence on the posterior distribution:

VSij=(∣yi−yj∣−1M∑m=1M∣xi(m)−xj(m)∣)2,VS_{ij} = (|y_i-y_j| - \frac{1}{M}\sum_{m=1}^{M} |x_i^{(m)}-x_j^{(m)}| )^2,

where yiy_i and yjy_j are actual field values at locations ii and jj, and xi(m)x_i^{(m)} and xj(m)x_j^{(m)} are the field values of the mm-th sample in the model posterior ensemble conditioned on the same history observations.

To test that BFF captures the joint spatial distribution, we compare the VS of spatially independent marginal distribution with that of the original joint posterior. We approximate the model's spatially independent marginal (i.e., product of pointwise marginal distributions) by randomly shuffling the sample index for each selected location. If the original joint posterior distribution has a VS lower than that of the marginal distribution, then we have evidence that BFF learns meaningful joint spatial distribution.

We summarize the results in the table below (4 displayed decimals). Please see Appendix J.6.1 for more detailed configuration of the evaluation.

NS moving sensorNS moving sensorNS moving sensorNS moving sensorNS uniformNS uniformNS uniformNS uniform
Spatial distanceOriginal VS ↓Shuffled VS ↓Rel. change after shuffle95% CI for relative changeOriginal VS ↓Shuffled VS ↓Rel. change after shuffle95% CI for relative change
1 pixel0.01420.18991233.99%[816.46%, 1761.40%]0.01310.0515+294.73%[160.98%, 439.62%]
4 pixels0.08800.162484.52%[53.40%, 123.36%]0.04670.0571+22.17%[11.98%, 30.03%]
16 pixels0.40020.40631.53%[0.006%, 3.19%]0.14370.1436−0.06%[−1.50%, 1.10%]
32 pixels0.55390.55470.14%[−0.68%, 0.86%]0.15680.1567−0.10%[−0.73%, 0.27%]

We observe that for the moving sensor or uniform sensor NS scenarios, the spatial joint posterior has a lower VS than the marginal one at spatial distances ≤16\le 16 pixels for moving sensors and ≤4\le 4 pixels for uniform sensors. We note that for uniform sensors, the sensor is placed with an spacing of roughly 128/8=16128 / 8 = 16 pixels, so the spatial correlation of the posterior uncertainty conditioned on the sensor observation would decay in this interval. In contrast, there is only one sensor in the moving sensor setting, which allows larger spatial regions for uncertainty.

AuthorsOpenReview ↗

Official Response to Reviewer STer (3/4)

(Response to Comment 3, continued)

Additionally, we compared the VS of BFF with the posterior modeling baselines in the NS uniform setting, averaged over the four pair distances (1, 4, 16, 32).

MethodOriginal VS ↓Shuffled VS ↓Relative change after shuffle
BFF0.090090.10224+13.49%
SDA0.053320.0560+5.04%
FlowDAS1.57692.7891+76.87%
EnSF0.58090.5827+0.31%
EnKF0.43040.4304−0.00%

Compared to the baselines, the original VS of BFF is the second best. We note that SDA is underconfident (Figure 4(b) in the revised manuscript) with a high miscalibration area ~0.43 (see Table 1 in the manuscript), and it makes predictions with less spatially-correlated fluctuations. This is reflected in its smaller relative VS change after shuffle. Meanwhile, FlowDAS has a worse VS absolute value. EnSF and EnKF demonstrate little relative VS increase after shuffling, suggesting their posteriors are not much more informative than the corresponding spatially independent marginals.

These results indicate that BFF's posterior captures reasonable spatially joint distribution. We have added these results in Appendix J.6.1 in the revised manuscript.

Comment 4 (W2 & Q4): Can the authors clarify the number of samples used in the KS experiments, the discrepency between Appendix J.1 and Table 2 latency benchmark results? Besides, can the authors re-evaluate the results with paired trajectory resampling, same sample count, and a new metric, the fair CRPS?

We thank the reviewer for raising these questions. Upon reviewing our configurations, we confirmed the discrepancies in the KS evaluation sample count and the latency benchmark in the original manuscript. We apologize for the confusion and address each point below:

  1. In the original manuscript, the sample count in KS varied across methods. EnSF and EnKF used an ensemble size of 64 for both updates and evaluation (as noted in the original Appendix I.1 and K.1), FlowDAS and SDA generated 16 samples for evaluation, and BFF used 100 samples in both original Table 1 and original Table 7. Upon careful inspection, we confirmed that this discrepancy stemmed from inconsistent configurations retained from preliminary experiments. To ensure a fair comparison, we re-evaluated all affected results using a unified sample size of 100 on KS and updated the manuscript accordingly. We confirm that our original central claims in the main text continue to hold.

  2. Previously, all non-deterministic methods used 16 samples for Burgers', 16 samples for NS, and 100 samples for Tokamak, except for EnKF on Burgers' (64 samples). We re-evaluated EnKF on Burgers with 16 samples and updated Table 1 in the manuscript. After the update, the only ranking change for BFF is that it moves from runner-up to best on spectrum deviation.

  3. We found that the relative L2 values in the original Appendix J.1 were also from an early experiment. To ensure a fair comparison under a consistet hardware environments, we re-ran the entire benchmark with comprehensive comparisons including all baseline methods from Table 1 across KS and NS. We updated Appendix J.1 in the manuscript, and the original claim on few-step acceleration still holds.

Based on the reviewers' suggestion, we aligned the sample count of all methods to 100 in KS. We also performed a paired trajectory bootstrap with 20,000 resamples to estimate the 95% confidence interval. We also added fair CRPS for the applicable methods (EnSF and EnKF ensemble members are coupled and thus excluded for fair CRPS).

Sample count MMRelative L2L_2 ↓Spectrum deviation ↓MA ↓CRPS ↓Fair CRPS ↓
EnKF1000.93010.66420.17501.0411—
EnSF1001.03130.76420.15221.2192—
FlowDAS1000.99000.78680.10741.12981.1195
SDA1000.91620.69130.07261.01061.0005
FNO1 (deterministic)0.97560.81461.00001.4866—
Transformer1 (deterministic)0.94780.80521.00001.4337—
BFF1000.87100.55490.20940.95880.9518
BFF delta vs best baseline [95% CI]—−0.0452 [−0.1316, +0.0412]−0.1092 [−0.3726, +0.1541]+0.1368 [+0.0693, +0.2043]−0.0518 [−0.1835, +0.0800]−0.0487 [−0.1805, +0.0831]

The results demonstrate that BFF is the best method on 4 out of 5 evaluated metrics, while the baselines are better on MA, which is consistent with the original Table 1.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response by Authors

We thank the reviewers for their detailed, constructive, and insightful comments. We are encouraged that the reviewers find our method novel (Reviewer VhJ9, 2RYr), of substantial contribution (Reviewer STer), and interesting (Reviewer VhJ9, 2RYr). Reviewer VhJ9 and 2RYr agree that our method is a potential alternative to particle-based filters.

Following the reviewers' comments, we did new experiments and edited the corresponding parts of the manuscript. We believe these updates have improved our work. The major updates are summarized below:

  1. We did additional ablation studies to isolate the benefits of BFF's weight-space TTT from other factors. We demonstrate that BFF benefits from the structural alignment with Bayesian filtering and can manipulate a large state space effectively. We compare our BFF with (1) a general neural-network parameterized update, (2) a generic parameterized loss, (3) a CFM conditioned on a dimension-matched GRU latent state, or (4) conditioned on all history information.

    These results are added to Appendix J.2, J.8, and J.10 in the revised manuscript, and also in our responses to weakness 1 from Reviewer STer, and the weakness & question from Reviewer VhJ9.

  2. We fixed discrepancies or insufficiently detailed descriptions in the evaluation description, including the normalization issue in NS, ensemble size and trajectory resampling in KS, and added a more comprehensive time benchmark on the same hardware.

    These results are updated in Section 4, Appendix J.1, and Appendix J.9 in the revised manuscript, with details also in our responses to questions 1 and 4 from Reviewer STer, and the weakness from Reviewer 2RYr.

  3. More evaluation metrics are added, including the variogram score that evaluates spatial dependence of the posterior distribution, fair CRPS in KS, and a diagnostic on Lorenz 63 that empirically shows the learned surrogate loss is positively correlated with the ideal filtering loss, and so do their gradients.

    These results are added to Appendix J.6.1, J.9, and J.11. Please see details in our responses to Reviewer STer questions 3 and 4, and Reviewer 2RYr questions 1 and4.

  4. We clarified that our theoretical contributions are two-fold, including the derivation of the weight-space Bayesian filtering rule that guides our method design, and a justification that (roughly speaking) our outer training loss controls the distributional error of the learned Bayesian filter.

    We have updated the Introduction in the revision to make our theoretical contribution clearer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.