acceptodds
Under review as a conference paper at ICLR 2027

Understanding Rollout Error in Graph World Models

Abstract

World models are increasingly used for planning by predicting future states under candidate actions. Many planning environments, however, are naturally graph-structured, with agents, tools, skills, routes, and dependencies connected through fixed or evolving relations. Understanding how prediction errors accumulate in such graph-structured rollouts is therefore critical for reliable long-horizon planning.In this work, we systematically study how rollout errors propagate in Graph World Models (GWMs). We analyze both fixed- and dynamic-edge regimes and derive topology-aware error bounds that characterize the roles of graph structure, learned dynamics, and evolving edges. We then validate these theoretical findings across synthetic graph topologies and heterogeneous agent-graph testbeds, showing that rollout error and planning regret increase with horizon and that dynamic-edge training is important when graph structure evolves. Together, these results characterize when GWMs remain reliable under long-horizon planning and when graph structure amplifies rollout errors. Motivated by these findings, we propose Error-Aware GWM, which combines spectral regularization, rollout consistency, and critical-node weighting, and improves long-horizon stability without sacrificing one-step accuracy. Experiments show that Error-Aware GWM consistently improves long-horizon rollout stability without sacrificing prediction accuracy.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph world models, rollout error, error propagation, dynamic graphs, spectral regularization, graph neural networks, model-based planning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_c79wOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how prediction errors accumulate in Graph World Models, which autoregressively roll out graph-structured states for planning. The authors analyze both fixed-edge and dynamic-edge regimes. For FE they derive a Graph Error Amplification Factor that cleanly separates topology effects from model effects, and for DE they derive a joint 2x2 node-edge error operator whose spectral radius governs coupled error growth. They also bound planning regret, showing it can grow super-linearly with horizon when the operator is expansive. Experiments on seven synthetic topologies and two agent-graph testbeds validate the theory, and they propose Error-Aware GWM with spectral regularization, rollout consistency, and critical-node weighting, which prevents divergence in vanilla GCN rollouts.

Critical strengths

1.The theory is clean and genuinely useful. The GEAF decomposition and the joint node-edge operator give an intuitive, testable account of why some graphs blow up under rollout, and the planning-regret connection makes it more than a toy analysis.

2.The empirical validation is systematic. Seven topologies spanning a wide spectral range, FE and DE regimes, perturbation and rewiring studies, plus heterogeneous agent testbeds, and the findings consistently support the predictions.

Critical weaknesses

1.The theoretical contribution is a fairly standard application of Lipschitz contraction arguments to graphs, so the bounds themselves are not very surprising to anyone familiar with model-based RL error analysis.

2.The proposed Error-Aware GWM is mostly a repackaging of known tricks. Spectral norm regularization, multi-step rollout training, and degree-based node weighting all exist in prior work, so the method's novelty is limited.

3.Several results are descriptive rather than confirmatory. GrowthSlope correlations look noisy, some tables have missing or inconsistent entries, and the real-world benchmarks show mixed results with DE-GWM sometimes worse than FE.

Questions for authors

Please see the Weaknesses.

6 comments
AuthorsOpenReview ↗

rebuttal to Reviewer c79w

1. Weakness 1 and Questions: Novelty of the theory

Appendix A acknowledges the existing foundations, including Asadi et al. [1] and Janner et al. [2]. Our contribution is a graph-specific error analysis. Section 3.3, Eqs. (3.5)–(3.8), separates node-to-node, edge-to-node, node-to-edge, and edge-to-edge sensitivities. This reveals how node–edge feedback affects error growth in DE rollouts, a pathway absent in FE prediction. Section 3.4 further connects the joint error bound to planning regret.

A scalar bound can describe the joint state but cannot directly explain how each pathway affects error growth. Our decomposition identifies these effects, explains the difference between FE and DE rollouts, and guides targeted feedback experiments. The contribution is therefore a graph-specific analysis of error propagation, not a new general contraction inequality.

2. Weakness 2 and Questions: Novelty of Error-Aware GWM

Section 4 acknowledges the existing components. Our contribution is to connect them directly to the graph error analysis. Each component targets a different source of long-term error: transition sensitivity, error from autoregressive inputs, and structurally weighted node error. Rather than simply combining existing losses, Error-Aware GWM uses the theoretical analysis to guide a unified training objective for stable and accurate rollouts.

Paper Table 1 reports median NodeMSE@32 of 0.3640.364 for spectral projection and 2.67×10−42.67\times10^{-4} for Error-Aware GWM. Rebuttal Table E2 (Reviewer zDcH, Response 2) confirms this comparison with corrected divergence counts. Rebuttal Table R1 (Reviewer F71V, Response 2) isolates the effect of each objective on the original backbone, while Rebuttal Table R4 (Reviewer zDcH, Response 5) shows that the benefits extend to MPNN. Together, these results support both the training design and its effectiveness beyond the original model.

3. Weakness 3 and Questions: Exploratory results, table inconsistencies, and mixed benchmarks

Appendix B.3 clearly labels GrowthSlope as exploratory, while Appendix D.3.3 reports OOD tests that do not remain significant after correction. Paper Tables 4–6 also retain mixed results on agent and real-world benchmarks. We will make these limits clear in the main text: a pooled correlation does not prove a general law of error growth, and our method does not outperform all baselines across all tasks.

We checked the inconsistent summaries against saved records and identified the errors. Rebuttal Table E2 corrects the two divergence counts. In Paper Table 2, the mean spectral excess ρ(B)−max⁡(LX,MA)\rho(B)-\max(L_X,M_A) across all 108 DE runs is approximately 77.077.0, rather than 155155. Agent records contain 50 test instances per testbed–seed cell, not 160 per testbed. We will correct these values and denominators. The reported accuracy medians and paired DE-training improvements remain unchanged.

We will also explain every missing result, including the MPNN out-of-memory failures reported in Appendix D.2.1. We will report attempted, completed, and divergent runs, paired-ratio calculations, and how non-finite errors are handled. Any analysis based only on successful runs will clearly state which runs were included.

Our code is public. We will release the full run manifest, raw metrics, failure records, splits, seeds, and aggregation scripts. We retain all unfavorable comparisons and report every planned experimental setting without selecting favorable seeds or datasets. The mixed results also clarify the scope of our method: Error-Aware GWM mainly targets long-term rollout stability, not universal improvements across all tasks.

[1] Kavosh Asadi, Dipendra Misra, and Michael L. Littman. Lipschitz Continuity in Model-based Reinforcement Learning. ICML, 2018.

[2] Michael Janner, Justin Fu, Marvin Zhang, and Sergey Levine. When to Trust Your Model: Model-Based Policy Optimization. NeurIPS, 2019.

As the rebuttal period ends tomorrow, please let us know if you have any remaining concerns. We would be grateful if you could revisit your assessment in light of our rebuttal and additional experiments. Thank you again for your review and consideration!

The rebuttal helped clarify the role of the node–edge decomposition and the reporting discrepancies. My concerns about the novelty of the training method and the strength of the empirical evidence remain, so I maintain my rating.

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

This paper investigates how prediction errors accumulate during long-horizon rollouts in Graph World Models (GWMs). The authors develop theoretical bounds for both fixed-edge and dynamic-edge settings, examining how graph topology, model parameters, and node-edge interactions influence prediction stability. Based on these findings, the paper proposes Error-Aware GWM, which combines spectral regularization, rollout consistency, and critical-node weighting to reduce error accumulation. Experiments across synthetic graph topologies, heterogeneous agent environments, and real-world graph benchmarks demonstrate promising improvements in rollout stability, although the gains are not consistent across all tasks. Overall, the paper addresses a relevant problem and offers useful insights into the relationship between graph structure and long-horizon prediction reliability.

Critical strengths
  1. I appreciate that the paper focuses on long-horizon error propagation rather than just one-step prediction accuracy. The joint treatment of node and edge errors is particularly valuable, as evolving graph structures introduce additional challenges for reliable planning. The proposed error-amplification framework provides a useful perspective on these interactions.

  2. The evaluation covers multiple graph topologies, fixed-edge and dynamic-edge settings, and different model architectures.

Critical weaknesses
  1. My main concern is the relationship between the theoretical error-amplification factor and the actual graph propagation operator. Section 2.2 introduces normalized adjacency for the fixed-edge GCN, whereas Lemma 1 uses the spectral radius of the unnormalized adjacency matrix. This distinction matters because normalization changes the operator's spectral properties and potentially the interpretation of topology-dependent error amplification. Additionally, the node-edge coupling inequality in Lemma 2 follows algebraically when the off-diagonal coupling terms are positive. Therefore, the reported statistical confirmation of this inequality does not independently establish the practical strength of node-edge feedback.

  2. Although Error-Aware GWM substantially improves stability relative to vanilla GCN, it does not consistently outperform stronger alternatives. For example, Table 2 shows that MPNN achieves lower median prediction error than Error-Aware GWM at the 32-step rollout horizon. The heterogeneous agent experiments also show limited improvements in task completion, while the real-world results remain mixed. On Cora, Error-Aware GWM achieves 75.0% accuracy compared with 80.6% for the standard GCN. Similarly, the dynamic-edge variants do not consistently outperform the fixed-edge model on Bitcoin-Alpha.

These results make the broader practical advantage less convincing. Additional discussion of when Error-Aware GWM is most beneficial would strengthen the paper.

Questions for authors
  1. Could the authors clarify why the theoretical error-amplification factor uses the spectral radius of the unnormalized adjacency matrix when the fixed-edge GCN formulation uses normalized adjacency? How does this affect the theoretical interpretation?

  2. Since the node-edge coupling inequality follows directly from the positivity of the coupling terms, what additional empirical evidence demonstrates that the measured coupling meaningfully predicts rollout behavior?

  3. Error-Aware GWM improves stability but does not consistently outperform MPNN or specialized graph models. Could the authors clarify which graph conditions or application settings benefit most from the proposed method?

10 comments
AuthorsOpenReview ↗

rebuttal to Reviewer zDcH

Thank you for the comments.

1. Weakness 1: The Operator and the Algebraic Coupling Inequality

Section 3.3, Eqs. (3.5)–(3.8), establishes a general error recursion using four transition sensitivities LX,LA,MX,MAL_X,L_A,M_X,M_A. This framework remains valid for normalized propagation when the constants match the actual model. We will correct Lemma 1 to account for the normalized operator, residual connections, and projection layers. These corrections preserve our main FE and DE theoretical results.

Appendix D.3.1 already explains that the spectral inequality follows mathematically when LAMX>0L_AM_X>0. We agree that this algebraic result alone is not independent empirical evidence. Our main empirical support comes from 36 paired comparisons in Section 5.5.1 (Paper Table 2), where DE training reduces rollout error by a combined factor of 11.011.0. This provides direct evidence of the benefit of modeling dynamic edges. Response 4 further tests node–edge feedback through controlled experiments.

2. Weakness 2: MPNN, Agent Tasks, Cora, and Bitcoin-Alpha

Our main contribution is improving long-horizon prediction accuracy and stability when rollout errors grow rapidly. Paper Table 1 compares several stabilization methods and shows that preventing divergence alone does not ensure accurate predictions. Rebuttal Table E2 provides stronger evidence: Error-Aware GWM achieves zero divergence across 200 runs and the lowest NodeMSE@32, reducing prediction error by over 1,360× compared with spectral projection. This demonstrates that our method improves both stability and long-horizon accuracy.

Rebuttal Table E2. Regularization comparison.

NodeMSE@32 and divergence counts across 200 attempted runs per method.

MethodNodeMSE@32 ↓DivergedAttempted
Vanilla GCN0.995100200
GCN + gradient clipping0.98781200
GCN + weight decay1.0915200
GCN + spectral projection0.36420200
Error-Aware GWM2.67×10−42.67\times10^{-4}0200

We also evaluate our method beyond unstable rollouts. Appendix D.4.4 (Paper Table 4) reports agent success rates near 0.420 across methods. Appendix D.5.1 (Paper Table 5) shows that Error-Aware achieves 75.0% accuracy on Cora, compared with 80.6% for GCN and 73.5% for FE-GWM. Appendix D.5.2 (Paper Table 6) reports Bitcoin-Alpha AP of 0.805 for Error-Aware DE-GWM and 0.820 for FE-GWM. MPNN also achieves lower error in some settings. These results do not weaken our main finding: Error-Aware GWM is designed to address long-horizon error amplification, where it delivers substantial gains in both accuracy and stability. We will clarify this focus and the method's practical scope in the revised paper.

3. Question 1: Why Use Raw Adjacency in the Theoretical Factor?

Raw adjacency was used to measure the graph structure in our GEAF diagnostic, not to represent the actual GCN propagation operator. The implemented undirected GCN uses normalized adjacency P(A)=D+−1/2(A+I)D+−1/2P(A)=D_{+}^{-1/2}(A+I)D_{+}^{-1/2}, where ∥P(A)∥2=1\lVert P(A)\rVert_2=1. We will correct Lemma 1 and Appendix H.1 to match this operator, including the residual and projection layers. Importantly, this correction preserves our main FE and DE error recursions and their theoretical conclusions. Raw spectral radius remains a useful diagnostic but does not directly bound normalized-GCN amplification. Rebuttal Table R1 (Reviewer F71V, Response 2) further shows that our Error-Aware objective improves long-horizon accuracy under normalized propagation.

AuthorsOpenReview ↗

rebuttal to Reviewer zDcH

4. Question 2: Does measured coupling predict rollout behavior?

Paper Table 2 shows a benefit from DE training, while Appendix D.3.2 and Paper Figures 14–15 show FE perturbation contraction. These observations distinguish the regimes but do not directly test whether measured coupling predicts long-term error growth.

Rebuttal Table R2 tests feedback and coupling prediction using existing DE checkpoints. Each node predictor remains autoregressive; only the edge input changes between predicted and true adjacency. Actions, initial states, test trajectories, and one-step errors are identical across conditions. Providing true edges lowers 32-step error in all 12 checkpoint pairs; panel (c) reports Error-Aware's median reduction by topology. This confirms that node–edge feedback affects long-term errors beyond the algebraic inequality. Panel (d) further examines whether measured one-step coupling predicts long-term error growth and perturbation gain.

Rebuttal Table R2. DE feedback intervention and coupling prediction.

Medians over three seeds; all 12 checkpoint pairs completed. NodeMSE@32 and GrowthSlope are shown in 10−310^{-3} units. Bold highlights the lower error within each model. Both inputs use the same checkpoints and trajectories; predicted edges reproduce the historical results.

(a) Scale-free

ModelEdge inputNodeMSE@32 ↓GrowthSlope
ActionNodePredicted18.110.9
ActionNodeTrue8.179.60
Error-AwarePredicted16.04.93
Error-AwareTrue4.14-0.0455

(b) Star

ModelEdge inputNodeMSE@32 ↓GrowthSlope
ActionNodePredicted16.03.17
ActionNodeTrue6.070.971
Error-AwarePredicted10.30.835
Error-AwareTrue2.18-1.17

(c) Error-Aware: reduction in median NodeMSE@32 with true edges

TopologyReduction
Scale-free74.1%
Star78.8%

(d) Predictive value of measured node–edge coupling

TopologySpearman ρ (GrowthSlope)Pearson r (GrowthSlope)Spearman ρ (Gain@32)
Scale-free0.8290.7640.817
Star0.8860.8310.853
Combined0.8460.7920.824

Together, these results show that node–edge feedback directly affects long-term errors. Panel (d) further shows that one-step coupling is strongly correlated with both long-term error growth and perturbation gain.

AuthorsOpenReview ↗

rebuttal to Reviewer zDcH

5. Question 3: When is Error-Aware GWM most useful?

Response

The strongest evidence comes from unstable long autoregressive rollouts, especially on scale-free and star graphs in Section 5.4.1, Paper Figure 4, and the rewiring study in Appendix D.2.1, Paper Figure 11. Rebuttal Table E2 shows that Error-Aware training improves long-term stability while maintaining prediction accuracy. Paper Table 2 separately shows the benefit of DE training when edges evolve.

These results suggest that Error-Aware GWM is most useful when long-term error growth is a major problem and changing edges affect node predictions. Graph density or degree alone does not determine its usefulness. Stable MPNNs and specialized models may still perform better, as shown by the Cora and Bitcoin-Alpha results.

Rebuttal Table R4 further tests whether Error-Aware training improves a stronger MPNN backbone. Both objectives use the same MPNN, DE edge head, parameter count, training epochs, data splits, and checkpoint selection rule. Only the Error-Aware losses differ. We include chain, scale-free, and star graphs and retain all seeds.

Rebuttal Table R4. Error-Aware training on the MPNN backbone.

Medians over three seeds; NodeMSE is shown in 10−310^{-3} units, while EdgeF1 is unscaled. Bold compares the two objectives within each topology. All 18 runs completed, with no divergence.

(a) Prediction accuracy

TopologyObjectiveNodeMSE@1 ↓NodeMSE@32 ↓EdgeF1@32 ↑
ChainStandard DE24.115.00.0343
ChainDE + Error-Aware17.013.00.0105
Scale-freeStandard DE38.416.90.107
Scale-freeDE + Error-Aware36.110.80.0296
StarStandard DE32.516.90.0105
StarDE + Error-Aware27.19.340.0312

(b) Reduction in median NodeMSE@32 relative to standard DE training

TopologyReduction
Chain13.2%
Scale-free36.1%
Star44.6%

Rebuttal Table R4 shows that Error-Aware training reduces 32-step node error in all nine seed pairs, with larger gains on scale-free (36.1%) and star (44.6%) than chain (13.2%). This supports its use when long-term rollout stability is important, even with a stronger backbone. EdgeF1 decreases on chain and scale-free but improves on star, showing that better node stability does not always mean better edge prediction.

The implementation is open source. We will report all configurations, seeds, aggregation rules, and failed or missing runs, including comparisons where our method performs worse.

Reviewer_F71VOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the rollout error in Graph World Models (GWMs) which autoregressively predict node features and optionally its adjacency. It separates a fixed-edge regime from a dynamic-edge regime. Theory parts claims the bound and empirical parts covers multiple synthetic topologies and baselines, FE and DE simulators, and synthetic agent-graph testbeds. The authors then proposes Error-Aware GWM (which ultilizeds spectral-norm penalty, multi-step rollout loss, and centrality-weighted loss), that avoids the divergence often seen with a vanilla GCN.

Critical strengths
  • This paper is clear over FE and DE strucutre, and well motivated over the compounding error in structured world models for agent and tool graphs. The FE/DE decomposition with an explicit node–edge operator is a nice framing.

  • The experiment results are comprehensive with agent task completion results and the GEAF-vs-degree correction comparison.

Critical weaknesses
  • The GEAF theory does not describe the model the paper uses, and the experiment results seem to contradict it. For example, the GCN transition in Eq. 2.4 uses the symmetrically normalized adjacency whose spectral norm is 1 regardless of topology while the proof of Lemma 1 uses the raw adjacency instead. If GEAF were the right quantity, errors should grow quickly on all reported topologies. Instead, the growth slopes in Table 1 are essentially flat, and Section 8.2 itself describes the fixed-edge dynamics as contractive. The complete graph has by far the largest GEAF while it has the lowest error and never diverges.

  • The benchmark shows no compounding error at all. In table 2 the errors go down with horizon for most methods and then level off at a small value which might suggest a contractive environment where compounding error isn’t really the issue. The only model that increases is vanilla GCN, and it already has large error at H=1 where this actually looks like a training problem rather than error accumulating over the rollout.

Questions for authors
  • Do the GCN experiments use the normalized or the raw adjacency?

  • Is there a setting where error clearly grows with horizon and GEAF predicts how fast?

3 comments
AuthorsOpenReview ↗

rebuttal to Reviewer F71V

Thank you for your comments.

1. Weakness 1: Normalization and the GEAF Bound

Lemma 1 provides an upper bound, not a claim that rollout errors must grow. Contractive trajectories are fully consistent with this bound, as shown in Appendix D.3.2 and Figures 14–15.

We agree that the bound must match the actual propagation operator. For Eq. (4.1), a valid constant is LX=Lσ∥P(A)∥2∥W1∥2∥W2∥2L_X=L_\sigma\lVert P(A)\rVert_2\lVert W_1\rVert_2\lVert W_2\rVert_2. The GCN uses normalized adjacency, for which ∥P(A)∥2=1\lVert P(A)\rVert_2=1, rather than raw adjacency with spectral radius ρ(A)\rho(A). The implemented residual GCN also requires the residual path and full projection matrices in the bound. We will correct Lemma 1 and Appendix H.1 accordingly.

These corrections preserve the FE and DE error recursions while making clear that the original GEAF is a topology diagnostic, not a direct bound on the implemented GCN. Its value alone does not guarantee error growth or predict its exact rate.

2. Weakness 2: Compounding Error versus Poor One-Step Fitting

Paper Figure 4 and Appendix D.1.3 (Paper Figure 9) show clear error growth within the same model's rollout. On the star graph, NodeMSE increases from 0.1130.113 at one step to 1.52×10351.52\times10^{35} at 32 steps (Rebuttal Table E1). This confirms that rollout errors can grow sharply with horizon. Meanwhile, stable baselines show that such divergence is not unavoidable.

Rebuttal Table E1. GCN rollout error — existing results.

NodeMSE, median over three seeds; source: Paper Figure 9.

HorizonScale-freeStar
10.5810.113
40.65512.0
162.77×1092.77\times10^{9}2.57×10152.57\times10^{15}
325.38×10235.38\times10^{23}1.52×10351.52\times10^{35}
48∞\infty∞\infty

However, these results alone cannot separate error amplification from poor one-step fitting or architectural differences. To address this, Rebuttal Table R1 compares four training objectives using the same Error-Aware backbone, data, training budget, and three seeds. All objectives use the same one-step validation criterion for checkpoint selection. Rebuttal Table R3 (Response 4) further tests error propagation using controlled perturbations.

Rebuttal Table R1. Same-backbone objective comparison.

Medians over three seeds. Units: NodeMSE@1, 10−310^{-3}; NodeMSE@32, 10−410^{-4}; GrowthSlope, 10−310^{-3}. Bold indicates the lowest error among the four new objectives.

(a) Scale-free

ObjectiveNodeMSE@1 ↓NodeMSE@32 ↓GrowthSlope
Prediction only1.592.455.82
Prediction + spectral1.592.294.31
Prediction + spectral + rollout1.572.123.15
Full Error-Aware1.701.651.12
Historical full †1.541.941.39

(b) Star

ObjectiveNodeMSE@1 ↓NodeMSE@32 ↓GrowthSlope
Prediction only1.191.954.72
Prediction + spectral1.211.783.21
Prediction + spectral + rollout1.181.621.85
Full Error-Aware1.221.250.42
Historical full †1.181.60-0.820

All 24 new runs completed without divergence. The four new objectives use the same validation checkpoint rule. † Historical results use final-epoch checkpoints and are excluded from best-value comparisons.

(c) Full Error-Aware: reduction in median NodeMSE@32 relative to prediction-only training

TopologyReduction
Scale-free32.65%
Star35.90%

Full Error-Aware reduces 32-step error by 32.65% on scale-free graphs and 35.90% on star graphs, despite having higher median one-step error. It also achieves the lowest GrowthSlope among the four new objectives on both topologies. These results show that the long-horizon gains cannot be explained simply by better one-step fitting. The controlled perturbation experiment in Table R3 further examines the role of error propagation.

Tables R1 and R4 use the implemented spectral penalty Rspec=∑ℓ(σ^ℓ−1)2R_{\mathrm{spec}}=\sum_\ell(\widehat{\sigma}_\ell-1)^2, estimated using four power iterations.

The weights are (λspec,λroll,λcrit)=(0.0100,0.500,0.100)(\lambda_{\mathrm{spec}},\lambda_{\mathrm{roll}},\lambda_{\mathrm{crit}})=(0.0100,0.500,0.100), with four-step rollouts. We will update Section 4 to match these implemented settings.

3. Question 1: Normalized or Raw Adjacency?

The GCN forward passes use normalized adjacency, as defined in Section 4, Eq. (4.1). The implementation also adds self-loops: P(A)=D+−1/2(A+I)D+−1/2P(A)=D_{+}^{-1/2}(A+I)D_{+}^{-1/2}, where D+D_{+} is the degree matrix of A+IA+I. Raw adjacency is used only for the reported GEAF diagnostic. We will clarify this distinction and update the theoretical bound to match the actual propagation operator, as explained in Response 1. These corrections preserve our main FE and DE error analysis.

AuthorsOpenReview ↗

rebuttal to Reviewer F71V

4. Question 2: Does GEAF Predict a Setting with Growing Error?

Paper Figures 4 and 9, summarized in Rebuttal Table E1, demonstrate clear error growth. Paper Figure 2 further shows a GEAF–GrowthSlope relationship, supporting GEAF as a useful diagnostic of error amplification. As stated in Appendix B.3, this relationship is exploratory rather than an exact prediction of growth rates.

We will refine Corollary 5 and Appendix H.5 to distinguish theoretical bounds from observed growth. An upper-bound recursion does not require the observed log-error slope to equal log⁡LX\log L_X. Moreover, the theoretical bound uses an error norm, while NodeMSE squares it, giving a factor-of-two difference in log slopes. Across the 12 R1 runs per topology, the operator-consistent bound has weak Pearson correlations with GrowthSlope (0.1040.104 on scale-free; −0.0682-0.0682 on star). Thus, our theory provides sensitivity bounds, while GEAF serves as an empirical diagnostic rather than an exact growth-rate predictor.

More importantly, Rebuttal Table R3 directly demonstrates the effectiveness of Error-Aware GWM in controlling error propagation. Each checkpoint is evaluated with and without a fixed initial perturbation under identical actions. At 32 steps, GCN amplifies perturbations in 138 of 420 trajectories (32.9%), while Error-Aware GWM shows zero amplification across all 420 trajectories, matching MPNN. Panel (b) further shows that GCN gains increase sharply from below one at the first step to extremely large values at 32 steps. These controlled results confirm that error amplification is a real problem and that our method effectively prevents it in the tested trajectories, independently of one-step fitting error.

Rebuttal Table R3. Controlled perturbation propagation.

Seven topologies × three seeds × 20 trajectories; all 63 planned cells completed, with 21 cells per method. Gains are medians of cell means; tangent errors are medians of within-cell medians. “Amplified” counts trajectories with gain above one at 32 steps; “Tested” gives the total trajectory count. Spearman correlation relates raw GEAF to gain.

(a) Comparison across methods

MethodGain@32AmplifiedTestedTangent error ↓Spearman ρ
GCN0.9900.9901384200.288%0.611
MPNN2.54×10−112.54\times10^{-11}04200.342%-0.430
Error-Aware GWM3.54×10−113.54\times10^{-11}04201.07%-0.474

(b) GCN median perturbation gain by topology

TopologyGain@1Gain@32
Scale-free0.8010.8011.05×10121.05\times10^{12}
Star0.5000.5002.59×10172.59\times10^{17}

The tangent calculation uses the full normalized transition, including residual and projection layers. Raw GEAF shows a positive correlation with GCN amplification (ρ=0.611\rho=0.611), supporting its value as a topology-based diagnostic. Its relationship with amplification varies across architectures, further supporting our operator-specific analysis. All outcomes are retained; gains near 10−1410^{-14} reach the numerical precision floor, where relative comparisons are less reliable.

Our code is open source. We will release the exact revision, configurations, seeds, checkpoint-selection rules, and all per-run metrics, including divergent outcomes.

As the rebuttal period ends tomorrow, please let us know if you have any remaining concerns. We would be grateful if you could revisit your assessment in light of our rebuttal and additional experiments. Thank you again for your review and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.