acceptodds
Under review as a conference paper at ICLR 2027

GraphBC: Graph-Embedded Bias Correction for Differentiable Hybrid Simulators under Model-Plant Mismatch

Abstract

Hybrid simulators combine mechanistic modules that are built on simplifying assumptions with data-driven modules learned from noisy or incomplete data. Local model–plant mismatches therefore compound through multi-stage computation, creating a dilemma: calibrating each module independently keeps intermediate states physically meaningful but leaves large terminal errors, whereas end-to-end training reduces terminal errors by letting modules compensate for one another's mismatch, at the cost of intermediate states and parameters losing their physical meaning. To address this dilemma, we propose Graph-embedded Bias Correction (GraphBC), which represents a hybrid simulator as a differentiable computation graph and embeds a learnable correction at an internal node, trained jointly with the simulator. We decompose simulation errors into propagated and local components and derive upper bounds along the error propagation. We then derive a gradient-alignment condition under which terminal supervision also improves intermediate states. Finally, we establish a turning-point criterion for correction placement in scaling-compatible chains. We evaluate GraphBC on two simulators driven by real-world data. On a freeway traffic simulator, it reduces terminal error by 29–43% relative to end-to-end training while shifting simulator parameters about 37% less, and gradient alignment identifies which intermediate states improve. On an industrial-scale ride-hailing simulator, it reduces terminal MAPE by 68% while preserving the physical meaning of intermediate states.

Primary area:
optimization
Keywords:
Differentiable hybrid simulators, Bias correction, Model-plant mismatch, End-to-end learning, Computation graph

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QWaMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a good question. Calibrating each module independently keeps intermediate states physically meaningful but leaves large terminal errors, whereas end-to-end training reduces terminal errors by letting moduels compensate for one another's mismatch. Authors treat simulator as differentiable computation graph and embeds a learnable correction at an internal node.

Critical strengths
  1. This is really a good question.
  2. Error propagation, gradient alignment and palcement analysis have clear meaning.
Critical weaknesses
  1. The error is only a Lipchitz expansion. Is there any mechanism that regulaize each module's dynamic?
  2. Under my understanding, this paper only adds a leanable small network as regulaization. This small network absorb mistakes. The mechanism seems too simple. The regualization only comes from errors, not from real dynamics.
Questions for authors

See Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Weakness 2

Dear Reviewer QWaM:

Thank you for your thoughtful comments. We acknowledge that our methodology was not clearly presented and further clarify it through a detailed comparison with dynamics regularization below.

  1. We agree that the propagation bound uses standard Lipschitz arguments. However, our framework and analysis extend beyond this bound in three ways:
  • Graph-based error decomposition: Separating local mismatch from propagated errors through module dependencies and downstream sensitivities.
  • Placement-aware bilevel formulation: Coupling correction placement selection with joint simulator–corrector training.
  • Internal–terminal gradient analysis: Establishing conditions under which terminal-loss updates also improve intermediate fidelity.

2. GraphBC introduces explicit model corrections rather than relying solely on dynamics regularization.

We thank the reviewer for raising this important point. Unlike dynamics regularization, which constrains model learning through additional loss terms, GraphBC introduces learnable corrections directly into the computational graph, extending the model's representational capacity . We further clarify the advantages of GraphBC in the following comparison:

DimensionDynamics regularizationGraphBCAdvantage of GraphBC
Correction mechanismLoss penaltyLearnable correctionDirect model correction
Mismatch handlingParameter adaptationResidual learningExtended expressiveness
Placement selectionPrespecifiedBilevel optimizationPerformance-driven placement
Internal–terminal trade-offPenalty-dependentGradient-alignment analysisExplicit improvement conditions
Error propagationNot characterizedGraph-based analysisLocal–propagated error distinction

The three key advantages of GraphBC are:

  • (a). Model expressiveness: Parameter adaptation vs. learnable correction. GraphBC extends the original model class through learnable residual corrections.
  • (b). Placement selection: Prespecified vs. optimized. GraphBC analytically selects correction placement through a bilevel optimization framework.
  • (c). Error characterization: Implicit vs. explicit. GraphBC characterizes local mismatch, downstream error propagation, and internal–terminal gradient relationships.

3. Experimental comparison with existing correction and regularization methods.

We added three representative baselines: APHYNITY (loss regularization), NeuralSim (simulator augmentation), and Solver-in-the-Loop (state correction), adapted to CTM under matched data splits, inputs, terminal supervision, and training budgets.

Results at 50% injected bias, with correction at C9C_9 where applicable, are summarized below. Mean RMSE equally weights terminal and internal RMSE.

MethodTerminal RMSEInternal RMSEMean RMSE
APHYNITY392.20429.21410.71
NeuralSim477.05448.57462.81
Solver-in-the-Loop277.60910.67594.14
GraphBC286.54449.80368.17

GraphBC achieves the best overall balance: terminal accuracy close to Solver-in-the-Loop (286.54 vs. 277.60), but with approximately 51% lower internal RMSE. It also achieves the lowest Mean RMSE, outperforming APHYNITY by 10.4%.

The expanded Table 2 confirms the lowest Mean RMSE across all six bias–placement settings. Implementation details and additional results are provided in the appendix.


4. Dynamics-embedded and dynamics-coupled correction in GraphBC

We clarify that GraphBC does not impose an explicit dynamics regularizer on every module. Instead, it embeds learnable corrections within the original simulator and trains them through downstream dynamical propagation. The mechanism consists of three aspects:

(a) Dynamics-embedded correction. GraphBC retains the original dynamical modules and introduces learnable residual corrections to compensate for mismatches beyond the original model's representational capacity:

x^i+1=fi(x^i;θi)+cϕi(x^i)\hat{\mathbf{x}}_{i+1} = f_i(\hat{\mathbf{x}}_i;\theta_i) + c_{\phi_i}(\hat{\mathbf{x}}_i)

The correction complements rather than replaces the original dynamics, extending the model's expressiveness while retaining its underlying computational structure.

(b) Dynamics-coupled training. Correction networks are trained using the terminal loss, with gradients backpropagated through the original downstream dynamical modules:

∂Lt∂ϕm=∂Lt∂x^T∂x^T∂x^m∂x^m∂ϕm\frac{\partial L_t}{\partial \phi_m} = \frac{\partial L_t}{\partial \hat{\mathbf{x}}_T} \frac{\partial \hat{\mathbf{x}}_T}{\partial \hat{\mathbf{x}}_m} \frac{\partial \hat{\mathbf{x}}_m}{\partial \phi_m}

(c) Graph-based characterization of correction–dynamics interactions. GraphBC uses module dependencies and sensitivities to distinguish local model mismatch from downstream propagated errors. Its gradient-alignment analysis further establishes conditions under which terminal-loss updates also improve intermediate-state fidelity. Together, these analyses explicitly characterize the interaction between learned corrections and the original dynamical system.

The supervision comes from prediction errors, but the correction updates are shaped by the downstream dynamics, rather than independent local error fitting.

Thanks for your reply. I acknowledge the comprehensive ablation study can address some of my concern. But I really cannot support a higher score for the novelty in ICLR, such a top-tier conference. I prefer to maintain my original score.

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

Overall, I think the paper describes an interesting point that is often overlooked in hybrid simulators, where low output error does not necessarily translate directly into low intermediate-state error. It can even go in the other direction, as learned components compensate for model mismatch to minimize the output error, making the intermediate states less physically meaningful.

Critical strengths
  • The paper studies the problem formally through a computation graph.

  • The authors provide a theoretical analysis of how errors propagate, when intermediate states can improve, and where the correction should be placed under restricted conditions. They also propose a simple solution using an auxiliary network that can be trained end-to-end with the simulator.

Critical weaknesses

Although the problem is interesting and the solution sounds reasonable and simple at first sight, after a closer look, I have some concerns:

  • The paper shows that whether output training also improves the intermediate states depends on the dot product between the gradients of the output loss and the intermediate loss (Prop. 4.1). When this is positive, a sufficiently small update also reduces the intermediate loss. Their solution is to add a correction network to absorb the mismatch and train it jointly with the simulator. However, this does not directly ensure a positive dot product, and the analysis does not explain how the correction makes this more likely. I feel the proposed solution is less directly linked to the theoretical analysis. This connection needs to be improved to strengthen the story of the manuscript.

  • Second, the authors analyze where the optimal correction node should be placed through Prop. 4.2, and identify a solution for a fixed-parameter, scaling-compatible scalar chain. However, in the traffic experiment, they do not use this criterion to choose the location, but rather directly use the validation error after training at all 13 candidate locations. Although the appendix validates the criterion in controlled cases, I still feel there is a gap between this theoretical analysis and what is done in practice.

  • A natural subsequent question is how to extend the solution to more than one correction location. Right now, each model uses only one correction location, and finding it requires a separate training run for each candidate. How would this work for larger graphs or when more than one location needs correction?

  • In the experiments, the comparisons are mainly with E2E training and their own ablations, such as Local and Correction Only. There is no direct comparison with other hybrid correction methods discussed in the related work, such as APHYNITY or NeuralSim. This raises concerns about whether the baselines are sufficient to show the benefit over existing approaches. Ideally, I would like to see these comparisons.

Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Weakness 1

Dear Reviewer usQN,

We thank the reviewer for this insightful comment. We agree that introducing a correction network does not guarantee positive gradient alignment. We clarify the connection between our theoretical analysis and GraphBC as follows.

(a) Internal correction changes terminal-driven parameter updates. As shown in Eq. (19):

B_{u,\eta_u}^{\top}A_{u\to t}^{\top} \nabla_{\widetilde{x}_t}\ell.$$ By modifying the simulated trajectory, correction changes the terminal residual and downstream sensitivities, thereby influencing the parameter-update direction and its alignment with intermediate-loss gradients. **(b) Positive gradient alignment enables intermediate improvement under sufficiently small terminal-loss updates.** Proposition 4.1 establishes this sufficient condition, which is numerically verified in Appendix E. With correction parameters and placement fixed, internal loss decreases at all 198 positively aligned checkpoints and increases at all 114 negatively aligned checkpoints after a terminal-gradient step of norm $10^{-6}$. **(c) GraphBC addresses this limitation through fidelity-aware placement optimization.** Rather than assuming positive gradient alignment, GraphBC explicitly incorporates both terminal and intermediate prediction errors into its upper-level placement objective: $$L_{\mathrm{place}}(\theta,\Phi;m) =\eta\left\lVert\widetilde{x}_t(\theta,\Phi;m)-y_t\right\rVert_2^2 +\sum_{u\in\mathcal{O}}\lambda_u\left\lVert\widetilde{x}_u(\theta,\Phi;m)-y_u\right\rVert_2^2.$$ Under linear assumptions, GraphBC analytically derives correction placement rather than training correction networks at every candidate location. We further evaluate all candidate placements experimentally to validate the analytical selection. **Revisions.** To strengthen the connection between theory and methodology, we have explicitly formulated GraphBC as a bilevel optimization framework, revised our contributions, expanded the comparison with prior methods, and clarified GraphBC's positioning in the manuscript.
AuthorsOpenReview ↗

Response to Weakness 2

We thank the reviewer for highlighting the gap between theoretical analysis and practical placement selection. We clarify their relationship as follows.

(1) Proposition 4.2 explains why correction placement matters. Under fixed parameters and scaling compatibility, all candidate placements match the terminal target but differ in intermediate fidelity:

Lplace(m+1)−Lplace(m)=−Δm.L_{\mathrm{place}}(m+1)-L_{\mathrm{place}}(m)=-\Delta_m.

The sign of Δm\Delta_m determines the preferred direction of placement, yielding an analytical optimum under the turning-point condition.

(2) The analytical result is verified in all 384 controlled CTM scenarios. Appendix E confirms agreement between analytical and recomputed optimal placement sets across three reference settings: all internal nodes, five observed internal nodes, and terminal-only supervision. Under terminal-only supervision, all placements achieve the same objective.

(3) Validation-based selection accounts for nonlinear dynamics and joint training. We acknowledge that Proposition 4.2 does not directly determine the optimum in the nonlinear CTM. Instead, validation terminal RMSE across 13 candidate locations selects C9C_9 for GraphBC and C13C_{13} for Correction Only. The test results further demonstrate the practical importance of placement:

Injected biasTerminal RMSE (C9C_9)Internal RMSE (C9C_9)Terminal RMSE (C13C_{13})Internal RMSE (C13C_{13})
50%286.54449.80290.60458.66
75%293.91450.14295.69472.12

At both bias levels, internal placement improves both terminal and intermediate accuracy.

In summary, Proposition 4.2 explains why terminal-equivalent placements can differ in intermediate fidelity, while validation-based selection identifies the preferred location after nonlinear joint training. We have clarified this distinction and the scope of the theoretical result in the revised manuscript.

AuthorsOpenReview ↗

Response to Weakness 3

We thank the reviewer for this valuable suggestion. We agree that extending GraphBC to multiple correction locations and larger computational graphs is an important direction for future work.

(1) The graph-based formulation can be extended to multiple correction locations. Placement can be represented by a set of correction nodes rather than a single node. The error-propagation analysis can then characterize their joint effects on downstream states, accounting for interactions between corrections.

(2) Graph sensitivity-guided sequential placement offers a potential scalable solution. Graph-based error sensitivities could be used to identify nodes with substantial influence on downstream prediction errors. Corrections could then be introduced sequentially, prioritizing locations with the greatest estimated improvement in the placement objective. Sensitivities would be updated after each selection to account for interactions with previously placed corrections.

(3) Multi-location optimization remains an important direction for future work. The present study focuses on single-location correction. Extending the theoretical guarantees and evaluating the computational efficiency of multi-location selection are beyond its current scope.

Reviewer_5bKBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The current work discusses the issue of "model–plant mismatch" in differentiable hybrid simulators, where localized modeling errors can propagate through the computational graph and ultimately accumulate into significant terminal errors. To address this, the authors propose the GraphBC framework, which introduces a learnable residual correction module at selected internal nodes. The manuscript makes several theoretical analyses of the model-plant mismatch issue and the GraphBC framework, and the empirical results demonstrate the advantages of jointly optimizing the added residual corrector with the main learning pipeline. However, the GraphBC method is heavily built on existing correction methods and the idea of residual correction. While the theoretical analysis offers some insights into the model-plant mismatch and GraphBC, some of these insights rely on stringent assumptions, and the guidance it provides for the real application of GraphBC is very limited.

Critical strengths
  • The paper addresses the trade-off issue between terminal prediction accuracy and the fidelity of intermediate states in h. Although this is a practically important question, it has not been systematically discussed before. To the best of the reviewer's knowledge, this paper is the first work to analyze this issue using a unified theoretical framework.
  • The current work has a solid experimental design. The reviewer especially enjoyed the ablations comparing Local, Correction Only, E2E Only, and GraphBC, which clearly disentangle the contributions of the corrector and joint training.
Critical weaknesses
  • The novelty of the GraphBC method itself is quite limited. Similar ideas for learning corrections have been used in many other works, as the authors suggest in Table 1. Although the authors tried to rephrase the idea in terms of a multi-module computation graph, they didn’t propose any new correction architecture or optimization method.
  • The theoretical analysis seems to be another major contribution of the current work. However, the actual value of the analysis is questionable. The correction placement analysis relies on some strong assumptions that are far from real-world applications. The author also suggests using validation-based selection for general nonlinear cases in Appendix D, which further weakens the method's significance.
Questions for authors
  1. Can the authors give a clearer definition of the GraphBC method, or can the authors highlight the differences between GraphBC and other baselines? The differences shown in Table 1 are more likely due to differences in the problem setup of each method, not to the methods themselves.
  2. Could the authors compare against suitable adaptations of other works, like the cited paper Solver-in-the-loop and NeuralSim, and include a purely neural predictor using comparable inputs, supervision, and training resources? The experiments compare GraphBC with Local, Correction Only, and E2E only, which do not sufficiently establish its advantages over existing hybrid modeling approaches. Where applicable, comparisons should report intermediate fidelity as well as terminal accuracy.
  3. In applications where intermediate ground truth is unavailable, how can we confirm that GraphBC actually addresses the local model mismatch and enhances the intermediate-state fidelity?
  4. GraphBC introduces additional trainable parameters from the correction network compared with E2E Only. Could the authors include capacity-matched baselines to distinguish the contributions of the residual correction, joint adaptation, and increased network capacity?
  5. Some figures are insufficiently self-contained. In particular, Figure 1 does not adequately explain the meanings of the node colors, empty and filled diamonds, observation symbols, and different arrows. It is also difficult to distinguish actual forward computation from the conceptual error-propagation illustration.
4 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

Dear Reviewer 5bKB,

We sincerely thank you for recognizing our work as the first to analyze the trade-off between terminal accuracy and intermediate-state fidelity within a unified theoretical framework. Below are point-by-point responses and revisions:

Response to W1 and Q1

We acknowledge that our original presentation did not sufficiently clarify the methodological novelty of GraphBC relative to prior work. We provide a clearer definition and a more detailed comparison below.

1. Definition of GraphBC

GraphBC is a unified optimization and analysis framework that couples correction placement optimization with graph-based error analysis in multi-module computational systems. It consists of two interconnected components:

(a). Bilevel optimization: GraphBC jointly considers correction placement and correction model training through a bilevel formulation:

\min_{m \in \mathcal{M}} \quad & L_{\mathrm{place}}(\theta_m^\star,\Phi_m^\star;m) \\\\ \text{s.t.} \quad & (\theta_m^\star,\Phi_m^\star) \in \arg\min_{\theta,\Phi} L_t(\theta,\Phi;m). \end{aligned}$$ The upper level selects the optimal correction placement, while the lower level trains the correction models for a given placement. **(b). Graph-based internal-terminal error analysis:** GraphBC uses the computational graph to explicitly characterize error propagatations across modules and to analyze the relationship between intermediate-state fidelity and terminal prediction accuracy under different correction placements. Thus, the methodological novelty lies in **coupling placement-aware bilevel optimization with graph-based error analysis**, rather than introducing a new correction network architecture. --- **2. Methodological Differences from Prior Work** We have revised and expanded Table 1 to provide a structured comparison of existing methods in terms of **problem settings and methodological approaches.** | **Method** | **Setting** | | | **Methodology** | | | | |---|---|---|---|---|---|---|---| | | Modules | Model | Placement | Correction formulation | Selection | Propagation | Internal vs. Terminal | | Post-hoc correction | Whole-system | Agnostic | Fixed | Calibration | Prespecified | — | — | | Discrepancy calibration | Whole-system | Agnostic | Fixed | Calibration | Prespecified | — | △ | | APHYNITY | Single | Mechanistic | Fixed | Learning | Prespecified | — | △ | | Solver-in-the-Loop | Single | Mechanistic | Fixed | Learning | Prespecified | — | — | | NeuralSim | Multiple | Mechanistic | Selectable | Learning | Connection discovery | — | — | | **GraphBC (ours)** | **Multiple** | **Hybrid** | **Selectable** | **Optimization** | **Optimization-based** | **Graph-based** | **✓** | Based on the above comparizon, the key differences are as follows: **(a). Placement selection: Prespecified placement (Solver-in-the-Loop, etc.) vs. optimization-based placement (GraphBC).** Unlike methods that rely on predefined correction locations, GraphBC explicitly optimizes correction placement through a bilevel formulation that couples placement selection with correction model training. **(b). Error propagation: Correction learning (NeuralSim, etc.) vs. graph-based error analysis (GraphBC).** Beyond learning correction models, GraphBC characterizes how correction placement influences error propagation across interconnected modules and affects both intermediate and terminal states. **(c). Internal–terminal trade-off: Terminal-focused optimization (Solver-in-the-Loop, etc.) vs. balanced fidelity (GraphBC).** GraphBC explicitly investigates the trade-off between terminal prediction accuracy and intermediate-state fidelity, demonstrating a more favorable balance between the two. --- **3. Summary of Methodological Contributions and Empirical Findings** Based on this comparison, we further clarify the three key contributions that distinguish GraphBC from prior work: **(a). Placement-aware bilevel optimization.** A novel bilevel framework that jointly optimizes correction placement and correction model training. **(b). Graph-based error-propagation analysis.** A graph-based characterization of placement-dependent error propagation across interconnected modules. **(c). Joint terminal and intermediate fidelity.** Experimental evidence of a more favorable balance between terminal prediction accuracy and intermediate-state fidelity compared with existing methods. We have revised Table 1 and the contribution statements in the Introduction, and added an explicit bilevel optimization formulation in Section 3, to clarify GraphBC’s methodological contributions and its distinctions from prior work.
AuthorsOpenReview ↗

Response to Weakness 2, Question 2 and Question 4

We thank the reviewers for these constructive suggestions. To address the concerns regarding baseline sufficiency, model capacity, and the practical value of GraphBC, we have added five baselines, covering purely neural modeling, capacity-controlled training, and established hybrid correction methods.

1. Additional baselines and experimental setup

(1) Neural and capacity-controlled baselines

  • Pure Neural Predictor: Replaces the CTM with a purely data-driven network to assess the value of retaining mechanistic dynamics.
  • Capacity-matched E2E Only: Increases the trainable capacity of E2E Only to match GraphBC, examining whether the improvements arise simply from additional parameters.

(2) Established hybrid modeling methods

  • APHYNITY: Augments physical dynamics with learnable corrections and penalizes correction magnitude, providing a comparison with regularized physics–data decomposition.
  • NeuralSim: Augments simulator components with neural networks to represent unmodeled dynamics, providing a comparison with neural simulator augmentation.
  • Solver-in-the-Loop: Trains neural state corrections through a differentiable solver, providing a comparison with dynamics-coupled correction learning.

All methods are adapted to the CTM setting using matched data splits, available inputs, terminal supervision, and training budgets. APHYNITY additionally includes a penalty on correction magnitude. Implementation details are provided in the appendix.

2. Experimental results

We organize the comparisons into three baseline groups and report both terminal and intermediate-state accuracy. Mean RMSE is the arithmetic average of Terminal and Internal RMSE.

The following results correspond to 50% injected bias, with correction placement at C9C_9 where applicable. Internal RMSE averages the RMSEs across five intermediate detectors. All errors are reported in veh/h.

GroupMethodTerminal RMSEInternal RMSEMean RMSE
Basic baselinePure Neural Predictor1867.852095.401981.63
Ablations and capacity controlsCorrection Only334.33520.62427.48
E2E Only448.99440.61444.80
Capacity-matched E2E Only455.23424.34439.79
Established hybrid methodsAPHYNITY392.20429.21410.71
NeuralSim477.05448.57462.81
Solver-in-the-Loop277.60910.67594.14
OursGraphBC286.54449.80368.17

3. Key findings

(1) GraphBC achieves the best overall terminal–internal accuracy trade-off. GraphBC attains terminal RMSE close to Solver-in-the-Loop (286.54 vs. 277.60), while reducing internal RMSE by 50.6%. Compared with APHYNITY, it reduces Mean RMSE by 10.4%. The expanded Table 2 further shows that GraphBC achieves the lowest Mean RMSE across all six evaluated bias–placement settings.

(2) GraphBC improves terminal accuracy over neural simulator augmentation while maintaining comparable intermediate fidelity. Compared with NeuralSim, GraphBC reduces Terminal RMSE by 39.9% (286.54 vs. 477.05), with nearly identical Internal RMSE (449.80 vs. 448.57). This supports the effectiveness of targeted mismatch correction in the evaluated CTM setting.

(3) GraphBC's gains extend beyond additional network capacity. Compared with Capacity-matched E2E Only, GraphBC reduces Mean RMSE by 16.3% (368.17 vs. 439.79). It also improves upon Correction Only by 13.9% (368.17 vs. 427.48), supporting the benefit of jointly adapting the simulator and correction network beyond residual correction alone.

Overall, these comparisons support the benefits of targeted mismatch correction and joint adaptation beyond simply increasing model capacity. They also demonstrate that GraphBC offers a more favorable balance between terminal accuracy and intermediate-state fidelity than the evaluated hybrid alternatives.

We have incorporated these additional baselines, experimental results, and implementation details into the revised manuscript and appendix.

AuthorsOpenReview ↗

Response to Question 3

We thank the reviewer for this important question. To assess intermediate-state fidelity without intermediate supervision, we introduce a leave-one-detector-out ablation, evaluating predictions against measurements excluded from both training and validation.

(1) Held-out detectors provide independent evidence of intermediate-state fidelity. We withhold each intermediate detector in turn and evaluate its prediction RMSE against the held-out measurements. The experiment uses 50% injected bias, retains the Local initialization, and places the source correction at C8C_8 for Correction Only and GraphBC. Solver-in-the-Loop retains its density-state correction. Baselines are selected by their terminal RMSE in the main comparison under 50% bias and placement at C9C_9.

(2) GraphBC achieves the lowest average prediction error at unseen intermediate locations. The leave-one-detector-out results are summarized below (RMSE; mean ± standard deviation where available).

DetectorGraphBC (all six)LocalCorrection OnlySolver-in-the-LoopGraphBC
C2C_2447.73 ± 0.59667.08667.09 ± 0.001560.13 ± 226.73473.23 ± 1.50
C4C_4527.52 ± 2.45527.24527.25 ± 0.00981.83 ± 227.67555.03 ± 3.83
C7C_7341.91 ± 1.38504.35502.37 ± 0.08892.65 ± 82.12351.46 ± 3.04
C9C_9332.74 ± 3.62741.83448.99 ± 1.871398.64 ± 1593.84386.56 ± 6.57
C12C_{12}433.91 ± 5.43480.72511.62 ± 10.951701.52 ± 1005.36556.77 ± 16.08
Mean416.76584.24531.461306.95464.61

(3) GraphBC improves intermediate fidelity without direct supervision at the evaluated locations. Compared with the selected baselines, GraphBC achieves the lowest average held-out RMSE, with the best results at C2C_2, C7C_7, and C9C_9. It also outperforms Solver-in-the-Loop at every held-out detector.

These results provide empirical evidence that GraphBC improves intermediate-state prediction accuracy even when the corresponding measurements are unavailable during training. While this does not directly isolate the reduction in local model mismatch, it demonstrates improved fidelity at independently evaluated intermediate locations.

Thank you for your responses and additional experiments. I appreciate the authors' efforts, and I believe that these clarifications and experimental results have helped improve the overall quality of the paper.

However, I have decided to maintain my original score after carefully considering the rebuttal, primarily for the following reasons:

  1. Concerns regarding novelty: My concerns about the novelty of the proposed approach have not been fully resolved. While the authors have provided further clarifications, I do not find them sufficiently convincing to change my original assessment. In particular, some of the criteria used to distinguish the proposed method from existing baselines appear to be subjective. I am not convinced that the authors of the compared baselines would agree with these characterizations.

  2. Unaddressed limitations of the theoretical analysis: The authors have not directly addressed my concerns regarding the limitations of the theoretical analysis, as raised in Weakness 2 of my original review. In my view, these limitations remain substantive and should be explicitly acknowledged and discussed. Given that the theoretical analysis is presented as one of the paper's main contributions, I do not believe that the current manuscript and rebuttal provides enough support for acceptance.

Overall, I find this work interesting and potentially valuable. However, I believe that the paper, in its current form, does not meet the standards for acceptance at ICLR yet. I would encourage the authors to further strengthen the manuscript, particularly by reconsidering how its main contributions are positioned before submitting a revised version in the future.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.