- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates the probabilistic forecasting modeling problem. The main methodology includes three parts. First, using history sequence to generate a deterministic predictive anhor. Then, add noise to the anchor. Finally, by graph-time velocity field and flow matching, map these initial states into future prediction. The experiments are comprehensive and results are significant.
Critical strengths
- The experiments are really comprehensive. This is worthy encouraging.
- This paper is well-orgainized and presented. It contains 40 pages and 32 tables.
Critical weaknesses
The experiments are really comprehensive, the limitations come from novelty and methodological contribution that are required by ICLR.
- All components are existing tools. Predictive prior, correlated source, GAT and flow matching are already utilized in many related fields. The loss includes three parts: prior, nll and CFM surrogate. It doesn't introduce any generative modeling theory. For example, authors can consider intorducing dynamics mechanisms to regularize the loss.
- According to the core of flow matching that transports noise to conditional distribution, this paper seems only use one-step residual, but not the ODE transport. The problem comes from the std from prior nll. Therefore, CRPS of 1/2/3 are 12.6744/12.6745/12.6747, indicating the learned speed is constant. Authors need to further check about this problem.
- Based on W2, the noise therefore is directly propagated.
Questions for authors
See Weaknesses.
6 comments
Part 2 of the Response (2/3)
W2. Is one-step decoding only residual prediction rather than Flow Matching?
Thank you for questioning whether the default decoder uses the learned flow field meaningfully. We clarify the distinction between path-state-conditioned training, the one-evaluation inference rule, and the prior’s role in setting the initial perturbation radius.
The training field is path-state-conditioned. For the linear conditional path,
the ideal unweighted and unclipped population field is
It depends on the current interpolated state, flow time, and context; the target displacement is therefore not a fixed history-only residual. Our implementation uses bounded endpoints and velocity targets, so we describe it as a bounded CFM objective rather than claiming the exact theory of the unclipped formulation (Appendix A.2).
The one-evaluation decoder is explicit. At inference, the default map is
where clips the updated state. We agree that this is algebraically a residual update. Its distinction lies in path-state-conditioned training and in applying the learned update to each source member. We describe it as a one-evaluation decoder, not an exact ODE solution or a second-order midpoint integrator (Section 4.4; Appendix A.4).
Prior scale sets the initial radius, not the final distribution. The prior-derived determines the amplitude of the initial graph-correlated perturbation. Each source member is then processed by the sample-state-conditioned field, so the final ensemble depends on both initialization and the learned member-specific update (Eqs. 3 and 8).
Similar CRPS does not imply a constant field. CRPS evaluates coordinate-wise marginal distributions and is relatively insensitive to how nodes and forecast steps are coupled. In our fixed-checkpoint diagnostic, four-evaluation RK4 improves GraphVS by 15.70%–20.75%, temporal variogram score by 54.86%–66.72%, and regional interval score by 5.29%–6.29% over one evaluation. This demonstrates that additional evaluations change measurable aspects of joint scenario organization, at approximately 3.94× the latency; the one-evaluation decoder remains the default (Appendix E.3).
The complete training scheme outperforms Score16. After twenty epochs, STGraphFM improves full-test CRPS by 2.50%, Energy Score by 2.53%, and GraphVS by 20.99% (Section 5.3; Table 3; Appendix C.2).
Part 3 of the Response (3/3)
W3. Does source noise simply pass through unchanged?
Thank you for identifying the possibility that the identity path could dominate the generated samples. We address this by distinguishing variation inherited from the source from additional corrections learned by the velocity field, using complementary retraining and fixed-weight interventions.
The identity path is not the learned update. The decoder has the form
so the source contributes an identity component. The key question is whether the velocity field learns meaningful, member-specific corrections beyond this inherited variation.
Shared retraining verifies sample-state dependence. The Shared control retains the source identity path, model capacity, graph, conditioning, and training objective, but replaces the current random-state input with the predictive background. This prevents the field from adapting its correction to individual source members. Compared with this control, STGraphFM improves calibrated CRPS by 3.35% and GraphVS by 21.48%; all six predefined trajectory-functional CRPS scores also improve (Appendix C.3; Table 12).
A fixed-weight intervention isolates learned corrections. To determine whether these gains come from the learned update rather than simply the source’s initial spread, we further compare native member-specific updates with shared-state and common-mean-velocity controls, using identical source draws and conditions. After removing the source identity contribution, native updates retain an across-member RMS dispersion of 0.526 ± 0.039, whereas both controls exhibit only floating-point residual dispersion (Appendix C.3; Figures 4–5).
These corrections improve joint forecasts. At identical native empirical marginals, native updates improve GraphVS by 8.14%, temporal variogram score by 13.09%, and regional interval score by 8.07%. Thus, the learned field contributes member-specific changes with measurable predictive value beyond the variation already present in the source (Figure 4; Appendix C.3).
The result concerns complete updates. These controls evaluate the full sample-state-conditioned mapping. Appendix E.1 separately examines graph-attention routing, so the demonstrated benefit is not attributed to dynamic routing probabilities alone.
Part 1 of the Response (1/3)
W1. What is the contribution beyond existing components?
Thank you for recognizing the breadth of the experiments and for asking us to clarify the methodological contribution. We have revised the Introduction and method presentation to connect the task-specific design with the controlled evidence supporting it.
The method separates three forecasting responsibilities. STGraphFM first estimates the predictable future background, then introduces graph-structured uncertainty around it, and finally learns how each perturbed future should evolve through a sample-state-conditioned graph–time field. The contribution is this coordinated adaptation for probabilistic spatiotemporal graph forecasting, rather than the invention of each underlying operator.
The objectives support these roles. SmoothL1 supervises the predictive center, Gaussian likelihood estimates its node–horizon uncertainty scale for source construction, and bounded velocity regression trains the sample-specific transport. The flow branch uses detached prior outputs, preserving the division between learning the background and updating future scenarios (Section 4.4; Appendix A.1).
Controlled studies support the design. Predictive centering improves mean CRPS by 12.90% with graph noise and 17.92% with IID noise. At identical predictive-center marginals, graph initialization improves GraphVS by 10.41% and regional interval score by 11.28%. Shared retraining and the fixed-weight intervention further support the predictive value of sample-state-dependent updates (Sections 4.1–4.3 and 5.3; Appendix C.1 and C.3).
Additional forecasting comparisons. We adapted GiFlow and FlowCast to the corresponding STGraphFM forecasting configurations. On KnowAirV2, STGraphFM achieves MAE/RMSE of 13.324/22.602, compared with 24.238/40.156 for our GiFlow-F adaptation. On WeatherBench T2M, STGraphFM achieves 1.272/2.164, compared with 1.389/2.331 for our FlowCast-WB adaptation, corresponding to 8.37% lower MAE and 7.19% lower RMSE (Appendices F.1 and F.3).
Physical penalties are domain-specific. We appreciate the suggestion. The benchmarks span traffic, air quality, and temperature, for which a common physical penalty would require additional domain-specific assumptions. We keep the current contribution focused on graph–time statistical generation and identify richer physical context as a natural extension.