acceptodds
Under review as a conference paper at ICLR 2027

Accelerated Sequential Flow Matching: A Bayesian Filtering Perspective

Abstract

Sequential probabilistic inference from streaming observations requires tracking distributions over future trajectories as new observations arrive. Although diffusion models can effectively capture complex distributions, their generation in streaming settings typically requires repeatedly transporting a non-informative source such as Gaussian noise to the target distribution. This noise-to-target transport incurs substantial inference latency and makes efficient high-fidelity sampling challenging. In this work, we introduce *Sequential Bayesian Flow Matching*, a framework inspired by Bayesian filtering. By learning a probability flow that transports the posterior distribution from one time step to the next time step conditioned on new observations, it mirrors the recursive structure of Bayesian belief updates. Crucially, as consecutive posteriors share common structure, the sequential flow only needs to learn the distributional update which enables substantially faster sampling than na\"ive resampling from scratch. Across scientific forecasting tasks spanning accelerator beam spill dynamics, fluid dynamics, and weather forecasting, as well as decision-making benchmarks, our method achieves performance competitive with full-step diffusion on distributional metrics while using 5-20 times fewer sampling steps, substantially reducing inference latency.

Primary area:
generative models
Keywords:
Diffusion, Flow Matching, Bayesian Filtering

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_7C5nOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

Modeling the changes of a dynamic system from streaming observations is challenging. There have been some classical filtering approaches and there are recent advancements in flow-based generative models, which are natural extensions to use. However, transporting from noise to the target is not easy, especially in a real-time streaming setting. To address this, this paper proposes Sequential Bayesian Flow Matching (SBFM), which learns a sequence of posteriors using flow-based models. This method has better performance than some baselines on a few scientific applications presented.

Critical strengths

None.

Critical weaknesses

1. Pretraining-finetuning framework

  • This approach raises some concerns about SBFM. With a standard training environment (which is described in line 257 of the main paper), I guess we can train SBFM but the performance is poor. We need an additional "pretraining" step to make the method better. Then, for example, for Figure 2-(b), aren't you just showing pretraining (rectified flow) vs pretraining + finetuning (SBFM) results? I'm not this is even a fair comparison. See questions below for further concerns.
Questions for authors
  • In real-world scenarios, ztz_t is often not independent of previous observations. Do you think your theoretical intuition will still hold even in this case?
  • When you train SBFM, do you need additional xt−1x_{t-1}, which is not used in the original conditional flow matching training?
  • If we train a flow map after pre-training or just a naive flow matching, do you think it will outperform your proposed method within a single step?
  • I'm confused about what exactly vθv_\theta will be when converged. If every loss goes to 00, what would be the optimal vector field? On the other hand, in practice, since the loss will never be 00, I wonder whether the loss accumulation is happening from a lot of places, including separate pre-training and fine-tuning phases (+ I also want to know how the "renoise" can be justified).
  • What are the sizes (number of parameters) for each baseline in the experiments? Also, did you use the same computational resources to train all of them?
  • Can Warm-start Diffusion/Wamr-start Rectified Flow be seen as SBFM methods as well? I think it still follows the method's architectural intuition but just not doing the fine-tuning.
  • Is "Pretrained Model" in line 363 the pretrained model for SBFM? If so, do you have results for SBFM + Diffusion as well for Table 3? Likewise, do you have results for SBFM + Rectified Flow for Table 4 and 5?
  • For experiments, the paper chooses N=20N=20 to represent the predictive distribution. This seems like a very small number to represent any distribution. Is there any reason you chose this number?
9 comments
AuthorsOpenReview ↗

Author Rebuttal: Regarding Pretraining-finetuning Framework

We appreciate the reviewer's time and efforts for providing the comments. Here is our response.

W1 (Pretraining-Finetuning Framework)

W1: "for Figure 2-(b), aren't you just showing pretraining (rectified flow) vs pretraining + finetuning (SBFM) results?"

We thank the reviewer for raising this concern. While we acknowledge that SBFM requires additional training effort, its fine-tuning objective primarily leverages bootstrapped data generated by the pretrained model to learn posterior-to-posterior flows, with the goal of improving inference-time sampling efficiency. Importantly, the improvement is not merely due to an extra finetuning compute alone, but our proposed SBFM objective.

To demonstrate this, in the following we conduct controlled experiments in which the same pretrained flow is further fine-tuned using conventional flow matching or flow map distillation under matched training computational budgets. The results show that SBFM consistently outperforms these alternatives at low NFE, highlighting the benefit of the proposed posterior-to-posterior flow learning objective. Furthermore, we show that the two-stage training procedure of SBFM can be unified into a single training stage by generating self-rollout data on the fly, eliminating the need for separate pretraining while preserving SBFM's superior performance.

1. Pretraining + conventional flow fine-tuning yields no further improvement

To separate the benefit of our formulation from the effect of extra fine-tuning, we continued fine-tuning the pretrained Rectified Flow model with standard noise-to-posterior flow matching, using the same data and compute budget as SBFM. This corresponds to setting the re-noising level to 1 in our pipeline. Standard fine-tuning gives only a marginal improvement over the pretrained model, while SBFM improves substantially (see Table C below). It implies the gains therefore come from the posterior-to-posterior objective rather than from additional training or computation.

Table C: Comparing SBFM and standard noise-to-posterior flow finetuning on a controlled synthetic state estimation task (see Appendix C.1 for concrete setup). The latter is implemented by SBFM with renoise level = 1 so that the source distribution becomes purely Gaussian.

MethodNFERMSE ↓
Pretrained Rectified Flow15.9
SBFM (renoise=1, effectively finetuning using naive flow matching)15.4
SBFM (renoise=0.4)13.2

2. Pretraining + flow map distillation underperforms SBFM

We also compare SBFM with two flow map distillation methods that, like SBFM, use a pretraining–fine-tuning pipeline: (a) AYF [1], closely related to MeanFlow, using its Eulerian Map Distillation variant; and (b) FreeFlow [2], a data-free method that mitigates teacher–data mismatch. Both compress noise-to-target transport into a single network evaluation, whereas SBFM accelerates inference by learning posterior-to-posterior transport. For a controlled comparison, both students are initialized from SBFM’s pretrained velocity networks and fine-tuned for the same number of epochs. As Tables D shows, AYF is unstable to optimize due to Jacobian computation, while FreeFlow improves few-step sampling but still underperforms SBFM.

Table D: Comparison with flow map distillation approaches (AYF and FreeFlow) and conventional Rectified Flow on beam spill and Burgers’ equation forecasting.

ModelNFEBeam (Energy Score ↓)Burgers’ Equation (Energy Score ↓)
Rectified Flow10.03160.1199
Rectified Flow20.02450.1082
Rectified Flow50.02240.1038
AYF10.18480.1361
AYF20.14420.1254
FreeFlow10.03040.1034
SBFM10.01650.1007
SBFM20.01590.1016

3. A single-stage SBFM training from scratch

We further show that we can trarin SBFM from scratch by a self-rollout training strategy. Specifically, instead of pretraining a conventional rectified flow model before fine-tuning SBFM, we initialize both models simultaneously and train them jointly. At each training step, the conventional rectified flow model is updated using the standard flow matching objective, while its current predictions are used to generate bootstrapped data on the fly for training SBFM with the proposed posterior-to-posterior flow objective. As shown in Table E, SBFM trained from scratch using this self-rollout strategy, with the same training dataset as the pretrained flow model, preserves the superior performance–efficiency trade-off.

Table E: W1 distance on synthetic random-walk dataset.

NFEPretrainedSBFM (Finetuning)SBFM (Self-rollout Training from Scratch)
16.8366.6876.483
23.6042.8463.155
32.4971.2760.884
41.9900.8460.463
51.8080.7460.245
AuthorsOpenReview ↗

Author Rebuttal: Regarding Questions For Authors

Here we answer the questions raised by the reviewer.

Q1: "ztz_t is often not independent of previous observations."

Our formulation does not assume ztz_t is independent of past observations, and the Bayesian filtering equations (1–2) require neither Markovian dynamics nor independent observations. By conditioning on the history z≤tz_{\leq t} and the previous posterior p(xt−1∣z≤t−1)p(x_{t-1}| z_{\leq t-1}), SBFM learns a posterior-to-posterior transport regardless of how ztz_t depends on previous variables.

Q2: "When you train SBFM, do you need additional x_{t-1}?"

SBFM needs samples xt−1x_{t-1} to learn posterior-to-posterior transport, rather than the noise-to-posterior transport of conventional flow matching. These samples come from model-based data augmentation: xt−1x_{t-1} is generated by the pretrained flow model, not drawn from an external dataset.

Q3: "If we train a flow map after pre-training or just a naive flow matching,..."

We find SBFM still outperforms naive flow fine-tuning and flow map distillation (Tables C, D).

Q4: "I'm confused about what exactly vθv_{\theta} will be when converged."

The optimal velocity field at the minimum of the fine-tuning objective transports the approximate previous pretrained posterior psource=ppre(xt−1∣z≤t−1)p_{source}=p_{\mathrm{pre}}(x_{t-1}| z_{\leq t-1}) (after renoising) to the true current posterior ptarget=p(xt∣z≤t)p_{target}=p(x_t|z_{\leq t}).

In practice, the pretrained posterior is only approximate and fine-tuning is imperfect. Errors can therefore accumulate over sequential updates, as the source distribution pgenerated(xt−1∣z≤t−1)p_{generated}(x_{t-1}|z_{\leq t-1}) generated at inference drifts from the one seen in training ppre(xt−1∣z≤t−1)p_{\mathrm{pre}}(x_{t-1}| z_{\leq t-1}). Re-noising reduces this mismatch by perturbing the previous posterior samples with Gaussian noise, which makes the model more robust to approximation errors. The re-noising ablation in Figure 6 of the manuscript shows that removing re-noising substantially degrades performance.

Q5: "What are the sizes for each baseline in the experiments?"

All methods share the same backbone (~2.5M, 300K, 2.5M, and 3.5M parameters for synthetic, beam-spill, Burgers', and weather, respectively).

Training-from-scratch methods use the same GPUs and number of epochs, though per-iteration cost varies by objective (e.g., MeanFlow requires Jacobian-vector products). SBFM uses the standard flow matching objective with a modified source distribution, generated by the pretrained model offline in advance, avoiding sampling overhead during fine-tuning.

Q6: "Can Warm-start Diffusion/Wamr-start Rectified Flow be seen as SBFM methods as well?"

From an inference-pipeline view, warm-start Diffusion and warm-start Rectified Flow are SBFM without fine-tuning. But the key difference is the learned transport. These methods are trained to transport Gaussian noise N(0,I)N(0, I) to the posterior ptp_t, but at inference they start from the (partially) noised previous posterior pt−1p_{t-1}. This mismatch between the source distributions at training and inference can greatly degrade sampling accuracy. SBFM instead learns the transport from pt−1p_{t-1} to ptp_t directly, so the training matches the inference-time initialization.

Q7: "Is "Pretrained Model" in line 363 the pretrained model for SBFM?"

Yes. The pretrained models in line 363 are the Diffusion and Rectified Flow models used to initialize SBFM, and both work with our framework. On Burgers’ equation, SBFM initialized from Diffusion beats the pretrained Diffusion model at the same NFE but trails SBFM initialized from Rectified Flow (Table F). We therefore reported the stronger-performing SBFM in our experiments."

Table F: Results of SBFM-Diffusion on Burgers’ equation forecasting.

MethodNFEL2 Energy Score↓RMSE↓
Diffusion100.1000.258
Rectified Flow50.1030.262
Diffusion10.1080.247
Rectified Flow10.1140.250
SBFM-Diffusion10.1050.242
SBFM-Rectified Flow10.1010.239

Q8: "For experiments, the paper chooses N=20 to represent the predictive distribution."

We thank the reviewer for the question. To check how sensitive the results are to particle count, we re-evaluated with N=100 particles. The numbers are overall robust with the N=20 results in the manuscript, and SBFM’s improvements remain (Table G).

Table G: Comparison of energy scores (lower the better) with N=20 and N=100 particles. Each entry reports N=20 -> N =100.

MethodNFEBeamBurgers’ equationWeather
Rectified Flow10.0302→0.03160.1140→0.12002.608→2.736
Rectified Flow20.0245→0.02450.1082→0.10802.503→2.502
Rectified Flow50.0231→0.02240.1030→0.10402.030→2.198
SBFM-Rectified Flow10.0179→0.01650.1010→0.10102.353→2.431
SBFM-Rectified Flow20.0170→0.01590.1016→0.10102.366→2.366

[1] Sabour et al. Align Your Flow: Scaling Continuous-Time Flow Map Distillation. NeurIPS 2025.

[2] Tong et al. Flow Map Distillation Without Data. CVPR 2026.

Thank you so much for the detailed rebuttal! I still have a few questions and I'll leave them below.

Questions to the answer of W1

  • I'm just a bit confused here. In the "pretraining" phase, are you training the exact same flow for both rectified flow and SBFM? When you see the training curve, is it converged? If so, what do you mean by "finetuning" more for the pretrained rectified flow? Can you explain more about how these training processes work, both for pre-training and fine-tuning, including the data and training objective? I think you are generating data from the pretrained flow and using that to fine-tune it... which I don't think is really a good idea (or fair)
  • How did you search for learning rates? Did you find optimal learning rates for each training method and compare them? Also is the reported number just a single seed with a fixed learning rate? If we were to validate things experimentally, I just want to see more completeness in the experimental setup so that we can really believe the numbers.
  • I'm also curious to see larger NFE comparisons
  • "As Tables D shows, AYF is unstable to optimize due to Jacobian computation" <- Is it impossible to run stable optimization for AYF in the Beam task? I wonder how bad the method is in terms of optimization. Can you share which implementation you follow for this?
  • Same for Table D. How confident are you in terms of training to optimality for the baselines? Since the number seems to be a single-seed training (or evaluation) value, it's not so convincing yet (I could be wrong so please provide more details if possible!).
  • Table E seems very interesting and good. However, (as in many on-policy methods) it seems like there could be a lot of difficulties in the initial phase of training. Were there any issues during training? Table E shows the synthetic case but does it also work without any error in more realistic tasks as well? Since the numbers look much stronger than the original proposed method, I'm very curious about this approach.

Questions to the answer of Qs

  • Q4: I think it's not clear yet. Can you give a clear formulation for the optimal vv such that when the loss is exactly 0? Also, you mentioned "Re-noising reduces this mismatch" but how exactly? Can you provide a formulation/theory for this? I don't quite get it. Define the exact mismatch and prove how re-noising can reduce it.
  • Q8: Thanks for the additional experiment! Can you also report the time it takes to sample for each method as well for different N? Also, can you also run different metrics so that you can capture better divergences between the distributions? Still N=100N=100 seems like a small number of particles to explain the whole distribution so I'd rather want to see other metrics as well.

Overall, once again, thank you so much for the additional experiments and efforts for the rebuttal. I am willing to adjust my score and confidence based on the answers to the questions above.

Reviewer_zD96OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Sequential Bayesian Flow Matching (SBFM), a method for efficient probabilistic inference in time-series setting where the underlying dynamics and observation models might not be known explicitly. Conventional flow matching approaches generate samples from the desired filtering distribution/posterior by pushing foward Gaussian noise at every time step. SBFM instead maintains particles approximating the previous posterior and learns to transport them to the updated posterior as new observations arrive. To ensure that the training pairs follow the correct conditional distributions, the authors use a pretrained flow-matching model to generate source samples for fine-tuning. Additionally they introduce a re-noising mechanism that perturbs previous particles with Gaussian noise to mitigate the mismatch between training and inference distributions as well as reduce sensitivity to accumulated approximation errors.

Critical strengths
  • SBFM tackles an important problem in an intuitive and simple manner.
  • The paper is clearly written and the approach well motivated by the existing literature on Bayesian filtering and flow matching.
  • The authors provide a broad evaluation on a wide range of tasks ranging from forecasting to planning and control.
  • The results demonstrate promising improvements in the trade-off between predictive quality and sampling cost, particularly for few-step inference.
Critical weaknesses
  • The energy score is defined using an L1 norm, which decomposes into scores of individual marginals and therefore cannot detect incorrect dependencies. Since the method aims to model whole distributions over entire trajectories an additional strictly proper score (using the L2 norm instead for example) would make the results even more convincing.

  • While the synthetic random-walk experiment compares the learned distributions against a known ground-truth distributions using Wasserstein distance, this evaluation is limited to a few updates, with errors averaged over time. A longer filtering experiment with noisy observations in a model that admits a tractable posteriors (e.g. a linear-Gaussian state-space model) would allow a more direct assessment of error accumulation. It might also offer some more insight on how the re-noising technique improves performance.

Questions for authors
  • Another experiment evaluating how posterior approximation errors evolve over successive filtering updates would be appreciated, ideally against tractable ground-truth posteriors.

  • Could the authors comment on the choice of the L1 norm in the energy score and consider reporting an additional metric sensitive to dependencies in the joint predictive distribution?

2 comments
AuthorsOpenReview ↗

Author Rebuttal

We thank the reviewer for recognizing that our work tackles an important problem in an intuitive way, is clearly written, and provides a broad evaluation. Here is our response to the comments.

W1&Q2 (Complementary Proper Scoring)

W1&Q2: "an additional strictly proper score (using the L2 norm instead for example) would make the results even more convincing"

We thank the reviewer for the constructive comment. As suggested, here we provide two complementary scores that capture temporal dependence. Given a ground truth trajectory xx and a generative model pθp_{\theta}, we compute (a) L2 energy score [1]: Ex^∼pθ∥x^−x∥2−12Ex^,x^′∼pθ∥x^−x^′∥2\mathbb{E}_ {\hat{x} \sim p_ {\theta}} \lVert \hat{x} - x \rVert_ 2 - \frac{1}{2} \mathbb{E}_ {\hat{x}, \hat{x}^{\prime} \sim p_ {\theta}} \lVert \hat{x} - \hat{x}^{\prime} \rVert_ 2; (b) Variogram Score [2]: ∑i<j(∣xi−xj∣−Ex^∼pθ∣x^i−x^j∣)2\sum_ {i \lt j} \left( \lvert x_ i - x_ j \rvert - \mathbb{E}_ {\hat{x} \sim p_ {\theta}} \lvert \hat{x}_ i - \hat{x}_ j \rvert \right)^2, where xix_i means the ii-th time step of trajectory xx. The Variogram Score (VS) explicitly evaluates temporal dependence by comparing pairwise differences between time steps in the observed and predicted trajectories. We find that the advantages of SBFM remain on those new metrics in low-NFE regime, as shown in Tables B.1-B.3 below.

Table B.1: Comparison of L2 energy score and Variogram score on beam spill forecasting.

ModelNFEL2 Energy ↓Temporal VS ↓
Rectified Flow10.05910.00550
Rectified Flow20.04220.00391
Rectified Flow50.03790.00332
SBFM-Rectified Flow10.02900.00243
SBFM-Rectified Flow20.02840.00232

Table B.2: Comparison of L2 energy score and Variogram score on burger’s equation.

ModelNFEL2 Energy ↓Temporal VS ↓
Rectified Flow10.1570.0182
Rectified Flow20.1440.0178
Rectified Flow50.1380.0180
SBFM-Rectified Flow10.1410.0171
SBFM-Rectified Flow20.1430.0176

Table B.3: Comparison of L2 energy score and Variogram score on weather forecasting.

ModelNFEL2 Energy ↓Temporal VS ↓
Rectified Flow13.5490.711
Rectified Flow23.2680.617
Rectified Flow52.8600.500
SBFM13.2010.555
SBFM23.0940.531

W2&Q1 (Error Evolution)

W2&Q1: "A longer filtering experiment with noisy observations in a model that admits a tractable posteriors (e.g. a linear-Gaussian state-space model) would allow a more direct assessment of error accumulation."

We thank the reviewer for the constructive suggestion. We agree that longer filtering experiments would provide further insights into error accumulation and the role of re-noising. While our work primarily focuses on demonstrating the effectiveness and computational efficiency of posterior-to-posterior transport, we acknowledge that distributional errors may accumulate over successive updates, potentially degrading long-term performance, as discussed in the Conclusion and Limitations section. A systematic investigation of long-term error propagation patterns and stability, and how potentially to mitigate it, remains an important direction for future work.

[1] Gneiting, Tilmann, and Adrian E. Raftery. "Strictly proper scoring rules, prediction, and estimation." Journal of the American statistical Association 102.477 (2007): 359-378.

[2] Scheuerer, Michael, and Thomas M. Hamill. "Variogram-based proper scoring rules for probabilistic forecasts of multivariate quantities." Monthly Weather Review 143.4 (2015): 1321-1334.

I thank the authors for their response and the additional experiments.

My concern about error accumulation and the role of re-noising remains open, however I appreciate the authors' acknowledgment of this limitation. Overall I will maintain my current score.

Reviewer_dKG7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Sequential Bayesian Flow Matching (SBFM) for probabilistic prediction with streaming observations. Instead of generating new predictions from Gaussian noise after every observation, SBFM learns to update samples from the previous predictive distribution. A pretrained model generates source samples using only previous observations, and a fine-tuned flow transforms these samples into updated predictions. The method also introduces controlled re-noising to mitigate accumulated errors. Experiments cover scientific forecasting, state estimation, planning, and control. The results demonstrate promising reductions in sampling steps, although the evaluation of complete trajectory distributions and the practical implications for the motivating real-time application require further clarification.

Critical strengths
  1. A useful formulation and training insight. The paper presents a well-motivated way to reuse previous predictions when new observations arrive. Section 3.4 identifies why directly pairing successive ground-truth states does not generally provide the source distribution available during inference. Constructing source samples with a pretrained model conditioned only on previous observations provides a meaningful rationale for the fine-tuning procedure beyond simply warm-starting an existing generator.

  2. Promising efficiency across multiple tasks. The evaluation covers several forecasting and control settings and includes multiple classes of acceleration baselines. SBFM achieves near-full-step forecasting performance on Burgers’ equation with one network evaluation and strong maze-planning results with three evaluations. The measured latency curves and ablations on re-noising and fine-tuning data provide additional evidence of its practical potential.

Critical weaknesses
  1. The main forecasting score may miss important errors in how a trajectory evolves over time. The “Energy Score” defined in Section 5.1 uses the L1L_1 distance. As written, this amounts to scoring each time point separately and adding the scores together. Consequently, two models can receive the same score even if they predict very different relationships between time points. For example, one model might predict that two consecutive days will both be cold or both be warm, while another predicts that the temperature will always switch between cold and warm. Both models can assign exactly the same temperature probabilities to each individual day, despite describing very different trajectories. This matters because the paper emphasizes modeling the distribution of complete future trajectories. The reported score and marginal coverage results do not establish whether these temporal relationships are captured correctly. Please confirm which distance is used in the implementation and supplement the evaluation with a metric that assesses the joint trajectory distribution, such as the Euclidean energy score, alongside a temporal-dependence diagnostic.

  2. The measured acceleration does not yet establish that the method meets the motivating application’s timing requirements. Section 5.1 motivates beam-spill forecasting with an approximately 1 ms feedback budget, but the fastest inference point in Figure 4 is approximately 9 ms. This is a useful relative speedup, yet it remains above the stated budget. The distinction matters because low-latency operation is central to the motivation for this application. Please clarify what the latency measurement includes and how the measured runtime relates to the intended deployment schedule. If the current implementation is not intended to meet the 1 ms budget, the practical claims should be framed as progress toward that requirement rather than evidence that it has been satisfied.

Questions for authors
  1. Does the implementation use L1L_1 or Euclidean distance for the reported Energy Score? If it uses L1L_1, do the main improvements remain when evaluated with metrics that also assess relationships across time points?

  2. What is included in the latency reported in Figure 4, particularly the number of particles, batch size, and preprocessing costs? Is the approximately 9 ms minimum the time required to produce one complete updated predictive ensemble?

  3. How is this runtime intended to fit the approximately 1 ms feedback budget discussed in Section 5.1? Does the proposed deployment use a different prediction-update frequency or another timing arrangement?

2 comments
AuthorsOpenReview ↗

Author Rebuttal

We thank the reviewer for the constructive comment. We particularly appreciate that the reviewer think this work provides a well-motivated and useful formulation and training insight. Here is our response to the concerns.

W1&Q1 (Supplementary Evaluation)

W1&Q1: “Please confirm which distance is used in the implementation and supplement the evaluation with a metric that assesses the joint trajectory distribution”

We thank the reviewer for the constructive comment. We use L1 Energy Score in the manuscript. Here we further evaluate two complementary scores that capture temporal dependence. Given a ground truth trajectory xx and a generative model pθp_{\theta}, we compute (a) L2 energy score [1]: Ex^∼pθ∥x^−x∥2−12Ex^,x^′∼pθ∥x^−x^′∥2\mathbb{E}_ {\hat{x} \sim p_ {\theta}} \lVert \hat{x} - x \rVert_ 2 - \frac{1}{2} \mathbb{E}_ {\hat{x}, \hat{x}^{\prime} \sim p_ {\theta}} \lVert \hat{x} - \hat{x}^{\prime} \rVert_ 2; (b) Variogram Score [2]: ∑i<j(∣xi−xj∣−Ex^∼pθ∣x^i−x^j∣)2\sum_ {i \lt j} \left( \lvert x_ i - x_ j \rvert - \mathbb{E}_ {\hat{x} \sim p_ {\theta}} \lvert \hat{x}_ i - \hat{x}_ j \rvert \right)^2, where xix_i means the ii-th time step of trajectory xx. The Variogram Score (VS) explicitly evaluates temporal dependence by comparing pairwise differences between time steps in the observed and predicted trajectories. We find that the advantages of SBFM remain on those new metrics in low-NFE regime, as shown in Tables A.1-A.3 below.

Table A.1: Comparison of L2 energy score and Variogram score on beam spill forecasting.

ModelNFEL2 Energy ↓Temporal VS ↓
Rectified Flow10.05910.00550
Rectified Flow20.04220.00391
Rectified Flow50.03790.00332
SBFM-Rectified Flow10.02900.00243
SBFM-Rectified Flow20.02840.00232

Table A.2: Comparison of L2 energy score and Variogram score on burger’s equation forecasting.

ModelNFEL2 Energy ↓Temporal VS ↓
Rectified Flow10.1570.0182
Rectified Flow20.1440.0178
Rectified Flow50.1380.0180
SBFM-Rectified Flow10.1410.0171
SBFM-Rectified Flow20.1430.0176

Table A.3: Comparison of L2 energy score and Variogram score on weather forecasting.

ModelNFEL2 Energy ↓Temporal VS ↓
Rectified Flow13.5490.711
Rectified Flow23.2680.617
Rectified Flow52.8600.500
SBFM13.2010.555
SBFM23.0940.531

W2&Q2&Q3 (Absolute Latency Constraint):

W2&Q2&Q3: "The measured acceleration does not yet establish that the method meets the motivating application’s timing requirements."

We thank the reviewer for the comment. We first clarify what the reported inference latency includes: it measures the average GPU execution time required to solve the flow ODE for one physical system (batch size = 1), whose predictive distribution is represented by (N=20) particles. These particles are processed in parallel on the GPU, and the reported latency measures the total time to generate all 20 trajectories, rather than the average time per trajectory. Thus, our measurement reflects the latency of a complete inference request for a single physical system, rather than per-particle throughput.

The total inference latency is approximately determined by the product of the number of function evaluations (NFE) and the latency per function evaluation. While we acknowledge that the current implementation does not strictly meet the approximately 1 ms feedback requirement, this work primarily focuses on algorithmic improvements, aiming to develop a general posterior-to-posterior flow framework that substantially reduces NFE while maintaining competitive distributional accuracy. The latency per function evaluation depends on factors such as the model architecture and computational hardware, and further latency reductions could be achieved through hardware–software co-design techniques, such as sparsity, quantization, and specialized FPGA or VLSI implementations, which are beyond the scope of this work. Finally, our experiments are conducted in a simulated environment, where predictions are evaluated without imposing real-time wall-clock constraints. We will clarify this distinction in the revised manuscript.

[1] Gneiting, Tilmann, and Adrian E. Raftery. "Strictly proper scoring rules, prediction, and estimation." Journal of the American statistical Association 102.477 (2007): 359-378.

[2] Scheuerer, Michael, and Thomas M. Hamill. "Variogram-based proper scoring rules for probabilistic forecasts of multivariate quantities." Monthly Weather Review 143.4 (2015): 1321-1334.

AuthorsOpenReview ↗

Would you mind letting us know if our response has addressed your concerns?

Dear Reviewer dKG7,

We would like to thank you again for the valuable feedback! Given that there are only three days left for discussion stage, we would really appreciate it if you could confirm whether our response has addressed your concerns.

To brieftly summerize our response, we:

  • regarding your concerns about trajectory-level evaluation, we provide two other proper scores as complemantary metrics. The results show that for those temporally-dependent metrics, our approch is still consistenly outperforming the baselines.

  • regarding your concerns about absolute latency contraints, we provide a detailed clarification on the latency computation protocal. We also emphasize that we focus on algorithmic improvements to reduce number of sampling steps, while samping latency per step depends on other factors like hardware and could be reduced through hardware–software co-design techniques.

If there is still any other issues, please kindly let us know and we are happy to follow up with you before the end of the discussion stage.

Thanks, Authors

Comments on the paper

1 comment

Dear Reviewers, the authors’ responses are now available. If you haven’t already, please review them and share your comments before the discussion period closes. Thank you for your time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.