acceptodds
Under review as a conference paper at ICLR 2027

BeliefGraph-JEPA: Structured Latent World Models for Action-Conditioned Time Series

Abstract

Action-conditioned time-series forecasting requires accounting for how future actions and exogenous forcings influence multiple targets through partially observed effects with different delays and persistence. Direct conditioning leaves the evolution and target-specific influence of these effects implicit in the predictor, while static relational graphs specify connections without tracking evolving effects. This motivates representing future-driver influence through structured latent states that evolve over the forecast horizon and route information to individual targets. We introduce BeliefGraph-JEPA, a structured latent world model that factorizes driver influence into typed latent-effect states. These states are rolled forward under future drivers and routed through a graph to target-specific nodes, forming the predictive base of a joint-embedding predictive architecture. A capacity-controlled residual supplements this base with direct driver information. On four multi-target clinical, agricultural, environmental, and industrial systems, the framework outperforms a range of pretrained and supervised known-future-covariate baselines. Matched controls isolate latent dynamics, future rollout, graph routing, and residual capacity; future rollout and graph-first residual routing improve forecasting across all four systems.

Primary area:
learning on time series and dynamical systems
Keywords:
action-conditioned forecasting, time-series forecasting, latent world models, structured latent dynamics, graph-based forecasting, joint-embedding predictive architectures

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_c1cVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates multivariate time-series forecasting in settings where future actions and exogenous inputs are known. In conventional time-series models, intermediate mechanisms through which such drivers affect the targets are often left implicit within the predictor. Alternatively, these intermediate mechanisms can be represented as a graph and modeled using a GNN, but such approaches do not explicitly capture how the effect of an action accumulates and decays over time. To address this issue, the paper proposes a method that represents driver effects as mechanism-specific latent states, rolls these states forward into the future, and propagates them through a graph to individual target variables. A key component is a graph structure specified in advance as a domain-informed prior. The method also introduces Future Rollout, which sequentially updates the latent states using known future drivers. Experiments on four different systems show that the proposed method achieves the best or near-best forecasting accuracy overall. The paper further conducts extensive ablation studies to analyze which of its many design choices are important. In particular, the result that relatively strong performance can still be obtained by retraining with a shuffled graph is noteworthy.

Critical strengths

One critical strength is that the paper proposes a novel approach to multivariate time-series forecasting that primarily leverages a graph structure specified in advance as a domain-informed prior, while also introducing Future Rollout to sequentially update latent states using known future drivers.

Another strength is the extensive ablation study. The proposed framework incorporates many design choices, and the authors make a substantial effort to disentangle their effects through a large number of ablation experiments.

Critical weaknesses

The paper introduces too many components, and the relationships among the problem formulation, the underlying hypotheses, and the individual technical components are not sufficiently organized. As a result, it is difficult to assess the necessity of each component and to identify the central technical contribution. Although the paper includes an extensive set of ablation studies, they do not sufficiently disentangle the main hypothesis of the proposed method from the contributions of its individual components, making it difficult to evaluate the novelty and technical significance of the overall framework.

In particular, while the sparse typed routing structure itself may be beneficial, the importance of the domain-informed semantic graph is not convincingly demonstrated. The shuffled-slot retraining achieves performance close to that of the domain graph, which suggests that the effectiveness of the domain graph, presumably one of the central technical contributions of the method, has not been sufficiently established.

Questions for authors

Could you revise the paper to more clearly disentangle the central hypothesis of the proposed method from the contributions of its individual components, so that readers can easily understand why each component is necessary and what constitutes the core contribution of the work?

2 comments
AuthorsOpenReview ↗

Response to Reviewer c1cV (Part 1)

We sincerely thank the reviewer for the thoughtful evaluation and constructive feedback. We especially appreciate the recognition of the novelty of combining a domain-informed typed relational prior with future latent-state rollout, as well as the extensive ablation study. We address the concerns below and clarify the central hypothesis and the roles of individual components in the revised manuscript (Introduction, Lines 47–49; Section 4, Lines 160–161).


Weakness 1: The framework contains many components, making it difficult to distinguish the central hypothesis from individual design choices and assess their necessity.

Response: Thank you for highlighting this issue. We would like to clarify that the central contribution is structured latent-effect world modeling with future rollout, rather than the collection of components or the exact semantic graph assignment in isolation. Our core hypothesis is that future drivers should update typed, persistent latent-effect states, which are explicitly rolled forward across the forecast horizon and used to predict multiple targets, instead of entering only as direct predictor covariates. The graph provides a structured driver–effect–target interface for this process, while graph-first residual routing (GR) makes the rolled representation the predictive base and confines direct driver information to a complementary residual path.

Our existing matched comparisons test distinct aspects of this hypothesis:

  • Persistent effects (RB vs. IE): Recurrent latent-effect states improve nMAE over instantaneous effects on all four multi-target datasets (Table 2).
  • Future evolution (FR vs. RB; FR-GR vs. RB-GR): Explicit future rollout improves nMAE in both matched comparisons across all four datasets (Table 2). The horizon-wise results (Figure 5) and later-action perturbations (Figure 4) further support the benefits of future rollout and its increased sensitivity to later future drivers.
  • Predictive pathway (FR-GR vs. FR; FLAT-GR vs. FLAT): The base–residual design improves both graph and unstructured readouts (Table 2). The routing/bottleneck controls (Table 3) further distinguish graph-only prediction, an unrestricted residual, and capacity-controlled correction. These results support GR as an integrated design; we do not claim that every subcomponent is independently indispensable.
  • Mechanistic capability (Figure 3 and Table 5): On matched synthetic dynamics, future rollout recovers the correct ordering of fast and slow persistences, and the full model exhibits responses on both wired driver–target pathways.

Together, these results support a progression from persistent latent effects → future rollout → structured predictive-base routing. The typed graph and residual mechanism serve distinct roles within this framework, with dataset-dependent contributions.


Weakness 2: The importance of the domain-informed semantic graph is not convincingly demonstrated, because shuffled-slot retraining performs close to the domain graph.

Response: Thank you for raising this concern. The domain graph achieves the lowest mean nMAE on both Greenhouse (0.1285 vs. 0.1309) and Wastewater (0.1245 vs. 0.1254) compared with shuffled-slot retraining (Table 4), although the differences are modest. The competitive shuffled-slot results suggest that the model can adapt its learned representations to an alternative routing assignment during retraining. Thus, we do not claim that one precise semantic assignment is indispensable.

Inference-time graph swapping offers complementary evidence: changing the routing assignment without retraining increases forecasting error in both directions (e.g., Wastewater nMAE rises from 0.1245 to 0.1610 when the domain-trained model uses the shuffled assignment). This indicates that the learned representations depend on the routing assignment used in training, rather than demonstrating that the domain semantics are uniquely correct. The learned fully connected graph also performs worse than the domain graph, particularly on Greenhouse (0.1357 vs. 0.1285), supporting the utility of the sparse typed prior relative to that alternative.

Our central contribution is structured latent-effect world modeling with future rollout, not the necessity of an exact domain-informed graph. The domain graph provides a relational inductive bias, while the gains from explicit graph routing and specific assignments vary by dataset (Tables 2 and 4).

AuthorsOpenReview ↗

Response to Reviewer c1cV (Part 2)

Question for Authors: Could you revise the paper to more clearly disentangle the central hypothesis from the contributions of individual components, so readers understand why each component is needed and what constitutes the core contribution?

Response: We appreciate this suggestion. The central hypothesis motivates two main technical components: (1) typed latent-effect dynamics with future rollout (Section 4.1), which models how driver-mediated effects evolve across the forecast horizon, and (2) graph-first residual routing (Section 4.2), which uses that structured trajectory as the predictive base while treating direct driver information as a capacity-controlled correction. The domain-informed graph supplies the relational prior for this pathway, rather than constituting a separate claim that each semantic assignment is necessary.

The manuscript already separates these components in Sections 4.1–4.2 and examines their roles through the matched ablations in Sections 5.1–5.3. To make their relationship to the central hypothesis more explicit, we have revised the Introduction (Lines 47–49) to state the central hypothesis directly, the opening of Section 4 (Lines 160–161) to clarify the complementary roles of latent-effect dynamics and graph-first residual routing, and Section 5.1 (Lines 308–312) to explicitly connect the ablation ladder to latent-effect persistence and future rollout.

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

This paper proposes a latent-effect-aware framework for action-conditioned time series forecasting. The key idea is to model the delayed and target-specific effects of future actions through structured latent states and graph-based routing. The authors evaluate the proposed framework on four datasets and conduct ablation studies to investigate the contributions of latent-effect dynamics, future rollout, graph structure, and residual routing.

Critical strengths

The experimental evaluation is comprehensive, particularly the ablation studies and additional experiments on synthetic dynamics and future-action perturbations, which provide further insights into the contributions of different components and the model's ability to capture latent dynamics and respond to future actions.

Critical weaknesses
  1. The improvements are relatively small on several datasets, which raise the concerns about whether they are statistically significant.
  2. As acknowledged by the authors, the learned latent-effect states may not correspond to meaningful mechanisms in real-world systems.
  3. The contribution of explicit graph structure is limited.
Questions for authors
  1. FLAT-GR and UR-GR achieve performance comparable to FR-GR on several datasets. Could the authors further explain under what conditions the explicit graph structure provides more substantial advantages?
  2. Given the relatively small improvements over simpler alternatives on several datasets, could the authors verify whether these improvements are statistically significant?
5 comments
AuthorsOpenReview ↗

Response to Reviewer HWCP (Part 1)

We sincerely thank the reviewer for the thoughtful evaluation and constructive feedback. We especially appreciate the recognition of our comprehensive ablations and the synthetic-dynamics and future-action perturbation experiments. We address the concerns below and report an additional paired-bootstrap analysis in the revised appendix.


Weakness 1: The improvements are relatively small on several datasets, raising concerns about statistical significance.

Response: Thank you for raising this concern. We have included a new Statistical Significance section in the Appendix L (Lines 1459–1511), with a paired-bootstrap analysis of the three-seed mean performance gaps (table below). For each comparison, the same resampled test units are used for both methods and all three training seeds, and we report percentile 95% confidence intervals from 10,000 replicates. We resample surgical cases on VitalDB, substations on PreDist, and within-series moving blocks on Greenhouse and Wastewater. We define Δ=m‾(baseline)−m‾(FR-GR)\Delta=\overline{m}(\text{baseline})-\overline{m}(\text{FR-GR}), so positive gaps favor FR-GR.

Importantly, FR-GR vs. RB-GR, which isolates future rollout under matched graph-first residual routing, has positive nMAE and nRMSE intervals excluding zero on all four datasets. The nMAE gains are 0.01178 on Greenhouse, 0.00807 on Wastewater, 0.00395 on PreDist, and 0.00030 on VitalDB. Thus, the benefit is not uniformly small, although the absolute improvement on VitalDB is modest.

Paired-bootstrap performance gaps (three-seed mean; 10,000 replicates). Δ=m‾(baseline)−m‾(FR-GR)\Delta=\overline{m}(\text{baseline})-\overline{m}(\text{FR-GR}); positive values favor FR-GR. Bold point estimates have percentile 95% intervals excluding zero.

DatasetBaselineΔ\Delta nMAE95% CIΔ\Delta nRMSE95% CI
GreenhouseRB-GR0.01178[0.01101, 0.01288]0.01581[0.01469, 0.01739]
FLAT-GR0.00585[0.00507, 0.00678]0.00764[0.00655, 0.00900]
UR-GR0.00456[0.00381, 0.00539]0.00617[0.00529, 0.00734]
WastewaterRB-GR0.00807[0.00658, 0.00996]0.00924[0.00653, 0.01247]
FLAT-GR0.00047[-0.00039, 0.00182]0.00006[-0.00146, 0.00230]
UR-GR0.00480[0.00363, 0.00633]0.00515[0.00342, 0.00739]
VitalDBRB-GR0.00030[0.00023, 0.00038]0.00050[0.00036, 0.00065]
FLAT-GR0.00003[-0.00005, 0.00011]0.00025[0.00012, 0.00038]
UR-GR0.00026[0.00014, 0.00038]0.00034[0.00013, 0.00054]
PreDist†RB-GR0.00395[0.00200, 0.00716]0.00638[0.00358, 0.01280]
FLAT-GR0.00151[-0.00116, 0.00262]0.00080[-0.00349, 0.00220]
UR-GR0.00102[-0.00249, 0.00216]0.00075[-0.00334, 0.00200]

Weakness 2: The learned latent-effect states may not correspond to meaningful mechanisms in real-world systems.

Response: We agree that predictive latent-effect representations should not be equated with identified physical or physiological mechanisms. Our objective is to explicitly model how the effects of future drivers evolve over time and influence different targets, rather than treating future drivers solely as direct predictor inputs. To this end, we represent driver-mediated dynamics through typed, persistent latent states, enabling the model to capture delayed, persistent, and target-specific effects. We do not claim that these learned representations uniquely recover ground-truth mechanisms from real-world observational trajectories.Our objective is to model driver-mediated dynamics through typed, persistent latent states, not to claim unique recovery of ground-truth mechanisms from real-world observational trajectories.

Section 5.4 and Appendix J already acknowledge this identifiability limitation. The matched synthetic experiment (Figure 3, Table 5) provides a more controlled capability check: the model recovers the correct fast/slow persistence ordering and exhibits responses along both wired driver–target pathways, with small off-path responses. We interpret these findings as evidence that the architecture can represent mechanism-specific dynamics under matched conditions, not as proof that every learned real-world state has a unique mechanistic interpretation.

AuthorsOpenReview ↗

Response to Reviewer HWCP (Part 2)

Weakness 3: The contribution of explicit graph structure is limited.

Response: We would like to clarify that our central contribution is structured latent-effect world modeling with future rollout, rather than the necessity of explicit graph message passing in every setting. We have further clarified this central contribution in the revised Introduction (Lines 47–49). The typed graph provides a relational prior that organizes driver–effect–target pathways, while latent-effect states evolve under future drivers. Graph-first residual routing makes the resulting structured trajectory the predictive base, with direct driver information entering through a capacity-controlled residual.

Thus, explicit GNN routing provides a structural inductive bias within the framework rather than being the sole source of its predictive gains. As discussed in Section 5.1, its additional benefits are dataset-dependent.


Question 1: FLAT-GR and UR-GR achieve performance comparable to FR-GR on several datasets. Under what conditions does explicit graph structure provide more substantial advantages?

Response: The clearest additional benefit of explicit GNN routing arises on Greenhouse. FLAT-GR retains family-specific latent-effect dynamics and future rollout but replaces GNN message passing with unstructured aggregation. Relative to FLAT-GR, FR-GR improves nMAE by 0.00585 (95% CI [0.00507, 0.00678]) and nRMSE by 0.00764 (95% CI [0.00655, 0.00900]). These intervals exclude zero; on Wastewater and PreDist both intervals include zero, while on VitalDB only the nRMSE interval excludes zero, with a small absolute gain.

Greenhouse involves multiple control and weather drivers affecting coupled climate targets. Target-specific routing of shared, heterogeneous effects may be more useful in such settings. UR-GR additionally replaces family-specific states with a shared state. FR-GR's intervals relative to UR-GR exclude zero for both metrics on Greenhouse, Wastewater, and VitalDB, but this comparison evaluates the combined contribution of latent-state factorization and graph readout, not GNN routing alone. These results are consistent with dataset-dependent benefits from explicit graph structure.


Question 2: Can the authors verify whether the relatively small improvements over simpler alternatives are statistically significant?

Response: Yes. We have added the paired-bootstrap analysis in Appendix L (Lines 1459–1511), summarized in the table above and documented its resampling protocol in revised Appendix L (Lines 1493–1511). The 95% intervals for FR-GR vs. RB-GR exclude zero on all four datasets for both metrics, supporting the benefit of future rollout under matched residual routing. In contrast, FR-GR vs. FLAT-GR has intervals excluding zero for both metrics only on Greenhouse (and for nRMSE alone on VitalDB); the remaining comparisons include zero. This distinction is consistent with our main claim: future latent-effect rollout provides the more consistent contribution, while explicit graph routing offers additional, dataset-dependent gains.

AuthorsOpenReview ↗

Looking forward to further discussion

Dear Reviewer HWCP,

Thank you for your time and effort in reviewing our paper. As the deadline for the rebuttal is approaching, we would like to kindly ask if you have any remaining questions; we are more than happy to address them.

The authors.

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

This paper proposes an architecture for forecasting several time series jointly when future actions are available. It maintains separate latent states for different kinds of accumulated input effects, advances those states under future inputs, and uses a typed graph to organize their influence on the predicted variables. The resulting graph representations and observation history form the main JEPA prediction, while a regularized correction can use future inputs directly. The contribution is evaluated on four multi-target domains, with additional glucose experiments and synthetic diagnostics.

Critical strengths

The approach is well motivated, and the breadth of the evaluation is a strength. In particular, the comparisons between future rollout and current-state prediction, including their residual-matched variants, provide useful evidence for advancing the latent states into the future. The flat-readout, shared-state, and shuffled-graph controls are also valuable. The authors are candid that the graph's additional benefits depend on the dataset and that good forecasts do not imply recovery of physical mechanisms.

Critical weaknesses

The two main weeknesses of the work are the following: an unresolved inconsistency in the predictor's stated temporal dependencies, and insufficiently convincing validation of the comparative results, particularly the synthetic baseline behavior and the documentation of external baselines. These concerns limit confidence in the conclusions drawn from an otherwise promising empirical study.

First, Section 4.1 describes the prediction z^t+h\hat{z}_{t+h} as depending on future inputs only through Dt:t+h−1D_{t:t+h-1}, but Appendix F, lines 975–982, specifies unmasked self-attention between horizon queries. Masking each query's direct access to future inputs does not necessarily preserve that restriction after queries exchange information.

For clarity, let sks_k denote an intermediate decoder state for horizon kk after incorporating graph or input information, rather than the learned query embedding QkQ_k. A subsequent self-attention head computes $\tilde{s}_h = \sum_{k=1}^{H}a_{hk} W_V s_k $

where ahka_{hk} are attention weights and WVW_V is the value projection. Without a causal query mask, weights for k>hk>h need not vanish. This permits the dependency path

Dt+k−1⟶Ht+kG⟶sk⟶s~h⟶z^t+h,k>h.D_{t+k-1} \longrightarrow H^G_{t+k} \longrightarrow s_k \longrightarrow \tilde{s}_h \longrightarrow \hat{z}_{t+h}, \qquad k>h.

The residual adapter permits an analogous path through later queries. Even if self-attention precedes cross-attention within a layer, information introduced in the first layer can be exchanged in the second layer of the reported decoder.

This is not evidence of future-target leakage, and full-horizon conditioning can be legitimate for the stated forecasting task. The issue is whether the complete predictor satisfies the prefix restriction used to describe it. The perturbation experiment in Section 5.5 still demonstrates input dependence, as the authors state, but its full-window error cannot distinguish a response in later predictions from changes to earlier ones. With history and exogenous inputs fixed, do predictions through horizon hh remain unchanged when actions change only after t+h−1t+h-1? Please clarify the attention ordering and masks and report this prefix-invariance check.

Second, the synthetic comparison does not yet clearly establish that the proposed dynamics explain the improvement over direct conditioning. Tables 24–25 report zero future-pulse responses for DCid\mathrm{DC}^{\mathrm{id}}, despite that model receiving future inputs and the data coming from a simple linear leaky integrator. For current-state models whose direct access to graph-covered future inputs is deliberately removed, the lack of a response is largely expected. The failure of DCid\mathrm{DC}^{\mathrm{id}} is less readily explained. Why does this baseline fail to learn a response? Checking its input access and convergence, and testing a simple supervised linear input-to-target model on the same synthetic data, would help distinguish an architectural advantage from an optimization or implementation issue in the controls. The single-seed evaluation and underestimated pulse magnitudes also limit the strength of this diagnostic. The authors appropriately present this as a matched-system capacity check rather than evidence of real-world identification. Nevertheless, the unexplained baseline failure limits what the comparison establishes even within that setting.

Finally, the external comparisons are difficult to assess from the reported training details. Appendix G gives the proposed method's configuration, but does not specify corresponding implementation, optimization, tuning, or checkpoint-selection details for TFT and DAG, or the adaptation and inference settings for the pretrained models. Appendix B also leaves the normalization ranges and dataset-specific missing-value thresholds unspecified. Giving methods the same future inputs is important, but does not establish comparable model selection. Could the authors provide these configurations, validation-selection procedures, and preprocessing constants? This matters because Table 2 shows competitive alternatives, including small graph-specific gains over FLAT-GR on Wastewater and VitalDB. Winning on every dataset is not a requirement for acceptance; the issue is whether the reported differences are sufficiently documented to assess their reliability and significance.

Questions for authors

Since Section 4.1 cites Schlichtkrull et al. (2018) and Appendix F specifies messages gWrhsrcg W_r h_{\mathrm{src}}, is the GNN a standard R-GCN or a custom variant? Please specify neighbor aggregation, normalization, and how action and exogenous-input information is processed separately and combined.

The authors should reconcile the temporal dependency description, explain the synthetic baseline failure, and provide sufficient details to assess the external comparisons. These are focused checks on the existing method and experiments, rather than a request for a substantially expanded benchmark suite.

6 comments
AuthorsOpenReview ↗

Response to Reviewer JVYM (Part 1)

We sincerely thank the reviewer for the careful assessment and constructive suggestions. We particularly appreciate the recognition of the motivation, breadth of evaluation, and matched comparisons isolating future rollout, current-state prediction, and residual routing. We have revised the manuscript to clarify temporal dependencies (Section 4.1, Lines 241–245), added a supervised synthetic control (Appendix K, Lines 1349–1394), and expanded baseline and preprocessing documentation (Appendices B and G, Lines 728–792 and 1114–1144). We respond point by point below.


Weakness 1: Temporal dependency of the predictor.

Response: We thank the reviewer for this observation. Our forecasting setting assumes that the full future-action sequence Dt:t+H−1D_{t:t+H-1} is specified at prediction time, as specified in Section 3, and the action-conditioned baselines follow the same future-input availability protocol. Therefore, joint horizon decoding with unmasked self-attention does not introduce future-target leakage or violate the stated forecasting setting.

We clarify that prefix causality applies to the latent-effect rollout, rather than the complete predictor. Specifically, the horizon-hh latent-effect state and its graph readout depend only on Dt:t+h−1D_{t:t+h-1}, whereas the subsequent joint decoder can integrate information across horizons through unmasked self-attention, as described in Appendix F. Accordingly, the complete predictor need not satisfy prefix invariance under the stated full-horizon known-action protocol. We have clarified this distinction in the revised Section 4.1 (Lines 241–245), including the complete predictor's dependence on the full future-driver sequence, consistent with the unmasked horizon-query self-attention specified in Appendix F.


Weakness 2: Synthetic baseline behavior and strength of the mechanism diagnostic.

Response: We appreciate the suggestion to compare against a simple supervised input-to-target model. We have added a prefix-linear ridge-regression control to Appendix K (Lines 1349–1394), trained and tested on the same matched-linear synthetic split as DC-id. For each forecast horizon, the control predicts from the last observed target and the corresponding future-driver prefix, with ridge penalty 10−410^{-4}.

MethodTest MAEFast-path pulseSlow-path pulseOff-path pulse
Ground truth—0.6000.2500.000
DC-id0.0630.0000.0000.000
Prefix-linear (new)0.0120.6020.2510.004
FR-GR-id0.0130.4300.1410.008

The prefix-linear control nearly recovers both ground-truth pulse magnitudes and achieves low forecasting error, even slightly below FR-GR-id on this matched linear system. This confirms that the matched synthetic input-to-target relationships are learnable through direct conditioning; the zero-pulse response observed for DC-id under this identification protocol is not a general representational limitation of direct conditioning.

These results clarify that the synthetic study is a matched-system capacity diagnostic, rather than evidence that structured rollout is uniquely capable of learning these dynamics. FR-GR-id recovers the correct persistence ordering and nonzero responses on both wired pathways, although their magnitudes remain underestimated. The new linear control and its interpretation are now included in the revised Appendix K (Lines 1349–1394).

AuthorsOpenReview ↗

Response to Reviewer JVYM (Part 2)

Weakness 3: Reproducibility and fairness of external baseline comparisons.

Response: We agree that the external comparisons should be sufficiently documented for readers to assess their reliability. We have expanded Appendices B (Lines 728–792) and Appendix G (Lines 1114–1144) with the previously unspecified preprocessing and baseline-selection details.

  • Preprocessing (Appendix B, Lines 728–792): We now report the dataset-specific missing-window thresholds and the physical lower/upper bounds used to normalize target and exogenous channels. We also specify the log-based [0, 1] normalization of VitalDB infusion rates, rather than min–max scaling, along with the 10-second resampling grid.
  • Supervised baselines (Appendix G, Lines 1116–1124 and 1135–1144): TFT and DAG use the same train/validation/test partitions and available future inputs as FR-GR. Their training uses Adam with learning rate 10−310^{-3} and validation MAE for checkpoint selection, evaluated once per epoch, with early stopping after 8 consecutive epochs without validation-MAE improvement. Their maximum epoch budgets are 50 on Wastewater and 30 on the other datasets, matching FR-GR's scheduled budgets. FR-GR itself runs its full budget and selects the lowest-validation-nMAE checkpoint. Appendix G records these distinct selection procedures and the remaining architecture/training settings. The baseline settings are fixed across seeds rather than selected by a hyperparameter search.
  • Pretrained baselines (Appendix G, Lines 1125–1130): We have also documented the zero-shot status, model checkpoints, future-covariate interfaces, and inference configurations for Chronos-2, Chronos-2-Cov, and TimesFM 2.5.

These additions document the preprocessing constants, future-input protocol, optimization settings, and distinct checkpoint-selection procedures used for the comparisons in Table 2.


Question for Authors: Is the GNN a standard R-GCN or a custom variant, and how are action and exogenous messages combined?

Response: Thank you for requesting this clarification. Our model uses a customized R-GCN-style relation-aware GNN, not the standard R-GCN of Schlichtkrull et al. (2018). At each layer, incoming edge messages are summed without degree normalization and the node update uses a residual connection followed by LayerNorm:

\qquad h_i \leftarrow \operatorname{LN}\left(h_i + \operatorname{Dropout}(\operatorname{GELU}(a_i))\right).$$ Here $W_r^{(\ell)}$ is a learned full relation-specific matrix (without R-GCN basis decomposition), and $g_{ji}=\sigma(\ell_{ji})$ is a learnable scalar edge gate initialized from the graph prior's edge confidence. Incoming messages are accumulated by sum aggregation, without degree normalization. Self-relations contribute to the message sum, while the residual preserves the untransformed node representation. Reverse edges use separate relation types and are only added for dynamic edges; exogenous nodes are source-only and their encoded features are restored after each layer. Action and exogenous sources use **separate** role adapters $A_A=I+\Delta A_A$ and $A_E=I+\Delta A_E$, both initialized at identity; effect and target sources use the identity mapping. **Action/exogenous fusion occurs before the GNN, not as two separate neighbor aggregations within it.** For each latent-effect family $f$, action and exogenous drives are aggregated separately and concatenated before the family-specific drive map, $u_{f,\tau}^{\mathrm{cont}}=C_f([q_{f,\tau}^A;q_{f,\tau}^E])$. The same role adapters are used to construct the two drive aggregates. The resulting drive updates the persistent effect state, which is injected into the graph before relation-aware message passing. We have clarified these implementation distinctions in the revised Appendix F (Lines 1043–1066), with a corresponding summary in Section 4.1 (Lines 226–230). The temporal-dependency, synthetic-control, and external-baseline concerns are addressed in the preceding responses and corresponding manuscript revisions.
AuthorsOpenReview ↗

Looking forward to further discussion

Dear Reviewer JVYM,

Thank you for your time and effort in reviewing our paper. As the deadline for the rebuttal is approaching, we would like to kindly ask if you have any remaining questions; we are more than happy to address them.

The authors.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers for their thoughtful feedback and constructive suggestions. We are particularly grateful to Reviewer c1cV and HWCP for positively reassessing our revised manuscript and increasing their score. We appreciate the recognition of our motivation, multi-system evaluation, and matched ablations. The comments helped us clarify the central contribution, strengthen the synthetic diagnostic and statistical analysis, and make the experimental protocol more reproducible. We have updated the manuscript, with changes highlighted in blue. Below we summarize the main edits.

Main Edits in the Manuscript:

  1. Clarified the central hypothesis and the roles of the components. We now state explicitly that the central contribution is structured latent-effect world modeling with future rollout, rather than a claim that a particular semantic graph is indispensable. The Introduction states the hypothesis (Lines 47–49), Section 4 explains the complementary roles of latent-effect evolution and graph-first residual routing (Lines 160–161), and Section 5.1 connects the ablation ladder to persistence and rollout (Lines 308–312).
  2. Clarified temporal dependencies and graph implementation. Section 4.1 distinguishes prefix-causal latent-effect rollout from joint prediction using the full available future-driver sequence (Lines 241–245). We also clarify action/exogenous drive fusion and the custom relation-aware GNN, including sum aggregation, role adapters, and relation-specific messages (Lines 226–230; Appendix F, Lines 1043–1066).
  3. Added a supervised synthetic control. A new prefix-linear ridge-regression baseline on the matched synthetic split shows that direct input-to-target conditioning can learn the impulse responses, placing the synthetic study's mechanism diagnostic in the appropriate context (Appendix K, Lines 1349–1394).
  4. Added statistical uncertainty estimates. We report paired-bootstrap 95% confidence intervals based on 10,000 replicates for FR-GR against RB-GR, FLAT-GR, and UR-GR, along with a dependence-aware resampling protocol (Appendix L, Lines 1459–1476 and 1493–1511). The intervals support a consistent benefit of future rollout under matched residual routing, while the added benefit of graph routing varies by dataset.
  5. Expanded preprocessing and baseline reproducibility details. We specify missing-window thresholds, validity/normalization ranges, and VitalDB infusion-rate scaling (Appendix B, Lines 728–792); and document supervised-baseline training and checkpoint selection, pretrained-model interfaces, and configurations (Appendix G, Lines 1113–1144).

Additional Clarifications During the Discussion:

We have also provided further clarifications in response to Reviewer JVYM's follow-up questions during the discussion. In particular, we clarified the checkpoint-selection criteria for the supervised baselines and their consistency with FR-GR and its matched ablations (Appendix G, Lines 1113–1124). We also verified DC-id's future-input access and training convergence, confirming that it receives future-driver inputs during both training and evaluation and that its optimization converges within the prescribed training schedule. While the specific cause of its zero pulse response remains unidentified, we have explicitly acknowledged this in Appendix K.3 (line 1384). We also incorporated the successful prefix-linear control into Section 5.4 (Lines 454-456). The corresponding manuscript sections have been updated accordingly.

We hope these additional clarifications address the remaining questions, and we remain happy to discuss any other concerns.

AuthorsOpenReview ↗

Response to Reviewer c1cV

Dear Reviewer c1cV,

Thanks for your thoughtful feedback and for reconsidering your evaluation! We sincerely appreciate your recognition of our clarifications and the revised manuscript.

The authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.