acceptodds
Under review as a conference paper at ICLR 2027

MuPS: Learning Multiple Physical Solution Branches from Unlabeled Outcomes

Abstract

Learning-based physical simulation is typically built around an input–output map: given a physical condition, a model predicts a solution, or samples one from a conditional distribution. This formulation becomes inadequate when the underlying physical problem is intrinsically multi-solution, where the correct prediction target is a condition-dependent set of admissible solutions. In such settings, a single predictor may average incompatible solutions, while a probabilistic model may require many samples to reveal distinct or rare outcomes. We identify this as a fundamental limitation of current learning-based physics models and formulate the problem as learning a set-valued solution map from branch-unlabeled input–solution observations. We introduce MuPS, which augments existing physics-learning architectures with a small collection of explicit solution hypotheses and condition-dependent activation. The hypotheses provide direct access to distinct solution modes, while activation determines which hypotheses are admissible for each condition, allowing the represented solution-set cardinality to vary across the input domain. We demonstrate that this formulation applies to Fourier neural operators, neural fields, and conditional diffusion models. Across three physical and synthetic benchmarks and three model families, MuPS substantially improves best-of-K prediction accuracy on multi-solution problems, often by orders of magnitude, while providing more sample-efficient access to distinct solutions. It remains effective under strongly imbalanced solution observations and can retain comparable performance on a standard single-solution problem.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Learning-based Physical Simulation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_nVqxOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

MuPS provides a novel formulation to directly learn condition-dependent sets of admissible physical solutions from branch-unlabeled input-solution observations, without requiring governing equations, physical objectives, or branch annotations during training. The method combines multiple-hypothesis learning with condition-dependent activation to select the admissible subset of candidate solutions for each physical condition. Its performance is assessed across three model families (Fourier neural operators, neural fields, and conditional diffusion models) and three benchmarks (reaction–diffusion, cloth simulation, and continuum robotics), primarily comparing against the corresponding unmodified models, with additional comparisons against multimodal prediction methods on a continuum-robot bifurcation experiment.

Critical strengths
  1. Broad applicability across model families. The formulation is largely architecture-agnostic and can be integrated into different physics-learning frameworks. The authors demonstrate its applicability to three model families, consistently improving multi-solution prediction accuracy across the studies physical benchmarks.

  2. Data-driven multi-solution learning. MuPS learns multiple solution hypotheses and their condition-dependent activation directly from branch-unlabeled input–solution observations, without requiring governing equations, physical objectives, solver structure, or branch annotations during training. This is particularly valuable for physical systems where the underlying equations or other problem-specific information may be unavailable or difficult to incorporate.

Critical weaknesses
  1. Limited comparison with existing multi-solution methods. Although learning multiple solutions is an established problem, the main experiments primarily compare MuPS against unmodified single-output models and conditional diffusion. While the improvements over these baselines are substantial, they provide limited evidence of how MuPS compares with existing approaches specifically designed to represent multiple solutions. The paper includes comparisons with established multimodal methods (Section 5.4) and alternative multiple-hypothesis parameterizations (Appendix A.8), but these are restricted to the continuum-robot benchmark. A broader evaluation against relevant existing multi-solution approaches would help better establish the relative merits of the proposed framework.

  2. Limited evaluation of varying solution multiplicity. Although MuPS is evaluated across three different physical systems, two benchmarks have a fixed number of admissible solutions across all conditions (four for reaction–diffusion and two for cloth). Only the continuum-robot benchmark evaluates changing solution multiplicity, with a transition between one and three solutions. Since adapting the number of predicted solutions to the physical condition is a central aspect of MuPS, evaluation on systems with more complex changes in multiplicity would provide stronger evidence of the approach's generality.

Questions for authors
  1. Evaluation of condition-dependent solution multiplicity. Two of the three benchmarks have constant solution multiplicity, while the continuum-robot benchmark exhibits a transition between one and three solutions. How do the authors expect MuPS to perform on systems with more complex changes in multiplicity, such as multiple bifurcations or several transitions in the number of admissible solutions?

  2. Estimating the number of admissible solutions. The activation model uses target solution counts estimated by clustering observations at the same and nearby conditions (Appendix A.3). How sensitive is this estimation to the choice of clustering threshold and the availability of repeated observations? In particular, how would the approach behave when only a few observations are available for each physical condition, or when distinct solution branches become very close near bifurcation points?

  3. Interpretation of branch-count results. Table 3 provides evidence that activation can recover the correct number of solution branches. However, some of the reported results are relatively trivial, while others are obtained on benchmarks where the number of admissible solutions is constant across physical conditions. This limits how much the table demonstrates about the effectiveness of condition-dependent activation.

    • Reaction-diffusion: For FNO and neural fields, the perfect branch-count accuracy without activation follows directly from setting the number of candidates K=4K=4 equal to the true number of solutions.

    • Cloth: The true number of branches is always two, and activation achieves 100% exact-match accuracy. However, it is unclear whether the same two hypotheses are retained for every condition or whether their identities vary with stiffness.

    • Robot: This is the most informative experiment because the true branch count (one or three) varies with the physical condition. It therefore provides meaningful evidence of condition-dependent cardinality prediction.

    Suggestions:

    3a. Report the candidate budget KK in Table 3 or its caption.

    3b. Include a condition-independent activation baseline that retains the same fixed subset of mm candidates for every physical condition. For instance, setting m=2m=2 means retaining the same two candidates across all conditions. With K=4K=4, there are six possible subsets of two candidates. The subset could be selected by minimizing symmetric set error on the validation set and then evaluated on the test set. Comparing its symmetric set error with that of condition-dependent activation would help determine whether the improvements in Table 2 arise primarily from consistently suppressing the same hypotheses or whether adapting hypothesis selection to the physical condition provides an additional measurable benefit.

  4. Clarity of Figure 2. The different panels are confusing, as they provide limited information about the problem setup, and the relationship between them is unclear. I suggest combining panels 1 and 2 to explain the problem setup as an input–output mapping, making it clear that the input is the physical condition aa and the output is one or several solutions uu, so that xx is not mistaken for the input. Panels 3 and 4 could also be combined to compare single-solution and multiple-solution approaches.

3 comments

Q1: Comparison with Existing Multi-Solution Methods

A: To further strengthen these comparisons, we have added evaluations against rMCL (Figure 22) and distance-based deduplication. Under the same 0.5 activation threshold, rMCL achieves 81.7% branch-count accuracy and a count MAE of 0.333 on the robot benchmark, compared with 88.3% and 0.200 for our method. On Cloth, distance-based deduplication yields 8.9–18.5× higher reconstruction error than our activation mechanism across all three backbones. Together, these comparisons provide additional evidence for the benefits of explicitly learning condition-dependent admissibility over alternative multimodal prediction and candidate-selection approaches.

Q2: Evaluation of More Complex Changes in Solution Multiplicity

A: We appreciate this suggestion and have introduced a new cantilever benchmark (Figure 23) to evaluate increasingly complex changes in solution multiplicity. The benchmark includes maximum stable-solution counts ranging from 1 to 4, with up to 20 branch-count transitions as the applied force varies. Our method accurately recovers the emergence and disappearance of solution branches at lower multiplicities and captures the overall bifurcation structure in more challenging settings, with some missed active regions at the highest multiplicity. These results extend our evaluation beyond the original continuum-robot benchmark and demonstrate the ability of our method to recover evolving solution sets across multiple bifurcations.

Q3: Robustness of Solution-Count Estimation

A: We would like to highlight an existing evaluation and a newly added experiment addressing the sensitivity of solution-count estimation.

  • Incomplete observations (existing Figure 21): Our ablation progressively removes training observations, showing that pooling neighboring conditions improves branch-count accuracy and set reconstruction compared with using only same-condition observations.
  • Threshold sensitivity (newly added Appendix A.8, Table 12): We compare automatic GMM and elbow threshold selection. Both achieve reasonable branch-count recovery, with the elbow method reaching 90% accuracy for FNO and neural fields, suggesting that our method does not require carefully tuned thresholds. Together, these evaluations provide complementary evidence of robustness to limited observations and threshold selection.

Q4: Branch-Count Recovery Near Bifurcations

A: We thank the reviewer for highlighting this challenging setting. We have added a new analysis in Section 5.3 that evaluates branch-count recovery near bifurcation points on the continuum robot benchmark. Considering a window of ±5 sampled conditions around each transition, we observe that branch-count exact-match accuracy decreases by approximately 5, 2, and 5 percentage points for FNO, neural fields, and diffusion, respectively, compared with regions away from bifurcations. These results demonstrate that our method maintains relatively high branch-count accuracy near transitions, while quantifying the additional difficulty of recovering admissible solutions when branches emerge or disappear.

Q5: Condition-Dependent Activation vs. Fixed-Subset Selection

A: We have added a fixed-subset baseline in Section 5.2 (Table 3), which selects a condition-independent subset by minimizing training-set symmetric set error and evaluates it on the held-out test set. Our activation achieves lower or equal symmetric set error across all three backbones on Cloth, with a particularly notable improvement for FNO (6.07 × 10⁻⁴ vs. 1.90 × 10⁻³). This demonstrates that even when solution multiplicity is constant, the appropriate learned hypotheses may vary across conditions. We further provide a breakdown by true solution multiplicity in Appendix A.8 (Table 13) on the robot benchmark. Our method outperforms fixed-subset selection in most multiplicity regimes, particularly for neural fields across all three branch counts, although some individual cases favor the fixed baseline. Together, these results demonstrate the benefit of adapting hypothesis selection to the physical condition rather than relying on a globally fixed subset.

Q6: Candidate Budget Clarification

A: We have explicitly reported the candidate budget K for each dataset and backbone in Table 4 (formerly Table 3) of the revised manuscript.

Q7: Clarity of Figure 2

A: We appreciate the suggestion and have revised Figure 2 to more clearly illustrate the problem formulation, distinguishing the physical condition as input from the potentially multiple admissible solutions as outputs.

I thank the authors for their detailed response. The updated paper addresses most of my original concerns and makes the results more informative.

I have three remaining observations:

1. Evaluation of rMCL. I have a concern about the fairness of the rMCL comparison. The authors note that rMCL normalizes its scores across hypotheses to sum to one. Applying an activation threshold of 0.5 to these scores makes it very difficult for rMCL to select more than one branch, and impossible to select two branches with scores higher than 0.5. This could explain why it always predicts one branch and achieves 81.7% exact-match accuracy, same as a trivial always-one baseline. Could the authors clarify how they selected the threshold? I think a validation-calibrated threshold or another selection strategy that allows rMCL to predict multiple active branches would make the comparison more meaningful.

2. Benefits of condition-dependent activation. The fixed-selection comparison is useful to understand the contribution of condition-dependent activation. Although the gains of adaptive selection are limited in some settings, it generally performs better than fixed selection when their results differ, especially on the robot benchmark. It would be useful to know how many hypotheses the fixed-selection baseline selects in each setting. This could be reported in the appendix.

3. Complexity of the benchmarks. The main benchmarks still have relatively limited variation in the number of solutions. For example, it appears 81.7% of the robot benchmark conditions have only one admissible solution, making branch-count accuracy less informative about performance in multi-solution regimes. The cantilever experiment is a useful addition, as it has more complex changes in solution multiplicity, including multiple bifurcations. However, the evaluation is mostly visual, and the method appears to miss some active regions at higher multiplicities. Could the authors also report quantitative metrics such as symmetric set error and branch-count accuracy for this benchmark? Exploring more challenging physical systems with complex bifurcation structures would be an important direction for future work.

I appreciate the authors' transparency in reporting results that show both the strengths and limitations of the proposed approach. Overall, I think the additional evidence and analyses have made the paper stronger, and I am inclined to upgrade my rating to clear accept.

We sincerely thank the reviewer for the encouraging feedback and thoughtful suggestions. We appreciate the recognition of our revisions and have conducted further evaluations to address the three remaining concerns.

Q1: Fair Comparison with rMCL

A: We thank the reviewer for pointing out the limitation of the fixed 0.5 threshold. While this threshold was initially applied consistently across methods, we agree that it restricts rMCL's ability to select multiple branches due to its normalized scores.

To provide a fairer comparison, we have calibrated the activation thresholds separately for both methods using the same protocol, allowing rMCL to select multiple branches. Our method achieves higher branch-count exact-match accuracy for FNO (95.0% vs. 91.7%) and conditional diffusion (91.7% vs. 81.7%), while matching rMCL for neural fields (93.3%). The corresponding count MAEs are 0.050 vs. 0.117, 0.083 vs. 0.333, and 0.100 vs. 0.083, respectively.

We have incorporated this evaluation into Appendix A.8 (page 32) of the revised manuscript.

Q2: Number of Hypotheses Selected by Fixed-Subset Activation

A: We appreciate this suggestion and have updated Section 5.2 (Table 3) to report the number of hypotheses retained by the fixed-subset baseline.

For reaction–diffusion, the fixed baseline retains all candidates (4 for FNO and neural fields, 6 for diffusion). For cloth, it selects 2 candidates across all three backbones. For the robot benchmark, it selects 2, 1, and 3 candidates for FNO, neural fields, and conditional diffusion, respectively.

Q3: Quantitative Evaluation of the Cantilever Benchmark

A: We thank the reviewer for this suggestion. We have added Table 14 to quantitatively evaluate the cantilever benchmark alongside Figure 23, reporting branch-count exact-match accuracy, count MAE, and symmetric set MSE across all four complexity settings.

For maximum solution multiplicities of 1–3, our method achieves 87.9–97.0% branch-count accuracy and consistently reduces symmetric set error compared with retaining all candidates. At the highest multiplicity (Nmax = 4), accuracy decreases to 60.6%, reflecting the difficulty of recovering active branches in more complex bifurcation structures. Nevertheless, activation reduces symmetric set error from 0.183 to 0.043, a 4.3× reduction.

These results complement the visualizations by quantifying both the benefits and limitations of our method as solution multiplicity increases. We agree that extending the evaluation to higher-dimensional physical systems with more complex bifurcations is an important direction for future work.

Again, we appreciate the reviewer's constructive feedback, which has helped strengthen the manuscript.

Reviewer_1P5ZOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies learning condition-dependent sets of physical solutions from input–solution observations without branch labels or governing-equation constraints. MuPS conditions a shared predictor on K learned solution features, trains candidates using best-match assignment and a separation loss, and subsequently learns an activation model with coverage, estimated-count, and duplicate-suppression objectives. The method is instantiated with FNOs, coordinate-based neural fields and conditional diffusion models. Experiments cover reaction–diffusion, cloth buckling and a simulated tendon-driven continuum robot, with additional studies of sampling budgets, rare branches, inverse inference, single-solution compatibility and hyperparameter sensitivity.

Critical strengths
  1. The paper addresses a meaningful limitation of single-output physical surrogates with a broadly applicable construction. Demonstrating the same hypothesis-and-activation formulation across three model families is useful, and Table 1 shows substantial improvements in best-of-K error.

  2. Set distance and branch-count metrics, the ten-seed imbalance experiment in Fig. 7, the matched-sample diffusion study, and the downstream inverse task provide complementary evidence. The appendix also examines K, separation, parameterization, activation weights, and incomplete observations, and acknowledges failure cases.

Critical weaknesses
  1. The methodological novelty beyond existing scored multiple-hypothesis learning is not sufficiently isolated. Section 5.4 and Table 10 compare useful alternatives, but do not establish the advantage of the proposed activation objective over these closer mechanisms.

  2. The reliability of condition-dependent admissibility remains insufficiently established. The observed failure cases in Tables 2 and 8, together with the absence of a direct analysis of activation in Appendix A.9, make it difficult to determine when the learned active set reliably preserves valid rare solutions.

Questions for authors
  1. Could the authors clarify the distinction between their support/admissibility scores and the learned scores in rMCL, and compare the activation objectives using the same candidate pool?

  2. What output-distance thresholds and normalizations are used to estimate MiM_i in Appendix A.3, and how were they selected using training/validation data?

  3. Could the authors specify the training/validation/test condition values and sample counts for cloth and robot, including whether all outcomes of one condition stay in the same split?

2 comments

Q1: Distinction from Existing Scored Multiple-Hypothesis Learning

A: Our key distinction is that we explicitly model the condition-dependent admissibility of each solution hypothesis, rather than its relative likelihood among candidates. The practical benefit is already demonstrated in our downstream inverse problem (Appendix A.5, Table 8), where activation reduces tendon-force recovery error from 5.06 N to 1.58 N.

To further clarify this distinction, we have added an rMCL comparison (Figure 22). Under the same 0.5 activation threshold, rMCL selects only one branch per condition, achieving 81.7% branch-count accuracy and 0.333 count MAE on the robot benchmark, compared with 88.3% and 0.200 for our method. Unlike rMCL's normalized scores under this threshold, our independent activation formulation supports simultaneous selection of multiple admissible solutions.

Q2: Reliability of Activation under Incomplete Observations

A: We appreciate this concern and highlight two complementary evaluations. Figure 7 (bottom) demonstrates near-complete rare-branch coverage even when rare solutions account for only 5% of post-bifurcation training examples. Additionally, Figure 21 presents an ablation on target solution-count estimation under missing observations, showing that pooling neighboring conditions improves activation reliability compared with using only same-condition observations.

These evaluations provide evidence of rare-branch preservation while clarifying the role of count estimation under incomplete observations.

Q3: Solution-Count Estimation and Clustering Thresholds

A: We have clarified in Appendix A.3 that solution-count estimation uses unnormalized pairwise L2 distances, with thresholds of 0.5 for Reaction-Diffusion and Cloth, and 0.36 for Robot. These were manually selected without systematic tuning or validation-based optimization.

We also added a threshold-sensitivity study in Appendix A.8 (Table 12) comparing automatic GMM and elbow selection. Both achieve reasonable branch-count recovery, with the elbow method reaching 90% accuracy for FNO and neural fields, demonstrating that carefully tuned thresholds are not required.

Q4: Data Splits and Condition-Level Separation

A: We have clarified in Appendix A.1 that all observations from the same physical condition remain in the same split. For cloth and robot, we use an 80/20 train/test split with a fixed random seed.

  • Cloth: 49 conditions (196 samples), split into 39 training conditions (156 samples) and 10 test conditions (40 samples).
  • Robot: 300 conditions (359 samples), split into 240 training conditions (279 samples) and 60 test conditions (80 samples).

These details are now included in the revised manuscript for reproducibility.

The authors' response provides clear and satisfactory clarifications. My concerns have been adequately addressed. I will maintain my positive rating and continue to recommend acceptance.

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

MuPS learns a condition-dependent set of physical solutions from input–solution observations without branch labels. A shared backbone conditioned on learned solution features produces multiple candidates, which specialize through best-match training and a separation regularizer. A separately trained activation network selects a subset using observation coverage, an estimated solution count, and duplicate suppression. The framework is demonstrated with neural fields, Fourier neural operators, and conditional diffusion on reaction–diffusion, cloth buckling, and continuum robotics. The practical motivation is sound, and the results show useful improvements in access to multiple outcomes under a fixed prediction budget.

Critical strengths
  1. The paper addresses a meaningful limitation of single-output surrogates in multistable systems. Explicitly addressable hypotheses can avoid averaging incompatible outcomes and reduce repeated sampling of common modes. The formulation can be incorporated into several existing model families.
  2. The paper includes relevant multimodal baselines in Figure 7 and alternative hypothesis parameterizations in Table 10. It also acknowledges dependence on data coverage, the absence of globally consistent branch identities, and the conditional scope of its theoretical analysis. The reported failure cases are useful for assessing the method.
Critical weaknesses
  1. The diffusion ablation is difficult to reconcile with the main results For reaction–diffusion, Table 1 reports a best-of-K MSE of 6.57e-5 for MuPS with conditional diffusion, and the Default row of Table 9 repeats this value. In Figure 18, however, the diffusion curve remains between roughly 0.1 and 0.2 for K = 1, 2, 4, and 8, several orders of magnitude higher. Table 4 specifies K = 6 as the default for this model, and that configuration is absent from the plot. The manuscript does not explain the large discrepancy or identify other configuration changes that would reconcile the results. This makes it difficult to assess how the reported accuracy gain depends on the hypothesis budget.

  2. The reference multiplicity is always four on reaction–diffusion and two on cloth in Table 3. Achieving 100% count accuracy on these datasets does not test whether the model tracks changes in the number of solutions. The main evidence for that capability comes from the continuum robot, where tendon tension varies and count accuracy is 88.3% for FNO and neural fields and 73.3% for diffusion. This is a useful demonstration, but the breadth of the cardinality-recovery claim is stronger than the diversity of settings that directly test it.

  3. The paper explicitly builds on established multiple-hypothesis and best-match learning. Its distinctive contribution therefore rests substantially on reconstructing a condition-dependent subset of the learned candidates. Tables 2–3 compare the proposed activation with retaining all K candidates. This isolates the benefit of filtering, but retaining every candidate is a limited baseline for set reconstruction: it does not establish the added value of the proposed coverage, count, and diversity objectives over a simpler pipeline of matching-derived labels, standard multi-label classification, and distance-based deduplication.

Questions for authors
  1. Could the authors explicitly state the hypothesis budget (K) used for each dataset and backbone in Tables 1–3? For reaction–diffusion, Tables 1 and 9 report a diffusion best-of-(K) MSE of 6.57e-5, whereas Figure 18 remains above approximately 0.1 for (K=1,2,4,8). Table 4 specifies a default of (K=6), which is absent from the figure. Are these results obtained under otherwise identical training and evaluation protocols? If so, what explains the large performance difference; if not, which configuration differences account for it?

  2. For reaction–diffusion with conditional diffusion, Table 3 reports 100% branch-count accuracy after activation, yet Table 2 shows that the symmetric set MSE increases from 1.16e-2 to 2.21e-2. What explains this deterioration? Could the authors clarify how activation changes the coverage and excess prediction components of the set error, and whether the retained candidates still represent all reference branches? This would help distinguish selecting the correct number of candidates from selecting a set that correctly covers the physical solutions.

3 comments

Q1: Experimental consistency between Table 1 and Figure 18

A: We thank the reviewer for catching this discrepancy. We identified that the diffusion ablation in Figure 18 was trained for only 1,000 epochs, compared with 5,000 epochs for the results in Table 1. We have rerun the ablation for 5,000 epochs and updated Figure 18 in the revised manuscript. The results are now comparable, providing a consistent evaluation across hypothesis budgets.

Q2: Clarification of hypothesis budgets

A: We appreciate the suggestion and have explicitly reported the hypothesis budget K for each dataset and backbone in Table 4 (formerly Table 3) of the revised manuscript, making the experimental configurations easier to identify and compare.

Q3: Generalization to more complex changes in solution multiplicity

A: To further evaluate this capability, we introduce a new benchmark with increasingly complex bifurcation structures (Figure 23), where stable solution branches repeatedly emerge and disappear as the applied force varies. We consider maximum solution multiplicities ranging from 1 to 4, with up to 20 branch-count transitions. Our method accurately recovers branch emergence and disappearance at lower multiplicities and continues to capture the overall bifurcation structure as complexity increases, with some missed active regions in the most challenging setting. These results provide new evidence that our formulation can recover evolving solution sets across multiple bifurcations, beyond the simpler transitions studied previously.

Q4: Comparison with stronger activation baselines

A: We appreciate this suggestion and have added comparisons with rMCL and distance-based deduplication to better isolate the contribution of our activation mechanism.

  • rMCL (Figure 22): Under the same 0.5 activation threshold, rMCL selects only one active branch per condition on the robot benchmark, achieving 81.7% branch-count accuracy and a count MAE of 0.333, compared with 88.3% and 0.200, respectively, for our method. Unlike rMCL's normalized scores under this threshold, our independent activation formulation allows multiple admissible solutions to be selected simultaneously.
  • Distance-based deduplication: We implement a baseline that removes near-duplicate candidates using calibrated output distances. On Cloth, it produces 8.9–18.5× higher reconstruction error than our activation mechanism across all three backbones and performs worse than retaining all K candidates. Unlike deduplication, our activation model uses observational evidence to distinguish admissible branches from distinct but unsupported predictions.

We also highlight two existing evaluations demonstrating the practical benefits of our approach:

  • Figure 7: Near-complete rare-branch coverage with only three queries, even when rare branches constitute just 5% of post-bifurcation training examples.
  • Appendix A.5 (Table 8): Activation reduces post-bifurcation tendon-force recovery error from 5.06 N to 1.58 N, demonstrating its benefit for downstream inverse problems. Together, these results highlight the advantages of explicitly learning condition-dependent admissibility for both efficient solution-set recovery and downstream physical inference.

Q5: Branch-count accuracy versus symmetric set reconstruction error

A: Correctly recovering the number of active branches does not necessarily minimize symmetric set error. In conditional diffusion, multiple candidates may approximate the same physical mode with different reconstruction errors. Activation can suppress a redundant candidate that nevertheless provides a closer match to a reference solution. Consequently, the reference-to-prediction coverage error may increase, while the excess-prediction component can change in either direction. This explains why branch-count accuracy can improve even when symmetric set error increases. We have clarified this trade-off between removing redundant hypotheses and preserving reconstruction accuracy in Section 7 of the revised manuscript.

Thank you for the detailed responses and additional experiments. The revised manuscript provides stronger evidence for the proposed approach. I will increase my score in light of these improvements.

I would like to flag one remaining reporting inconsistency between Tables 2 and 3. For conditional diffusion on Cloth:

  • Without activation: Table 2 reports 3.64e-1, whereas Table 3 reports 3.64e-2.
  • With activation / Ours: Table 2 reports 4.12e-2, whereas Table 3 reports 4.12e-4.

Both tables appear to report symmetric set MSE for the same configurations. Please double check these values, or clarify any differences in the evaluation settings that account for the discrepancy. I would also encourage the authors to cross-check corresponding results across the other tables to ensure that all reported numbers are consistent throughout the revised manuscript.

We sincerely thank the reviewer for the encouraging feedback and careful review. We apologize for the inconsistency and have corrected Table 3 to match Table 2. We greatly appreciate the reviewer's constructive suggestions and attention to detail.

Comments on the paper

1 comment

General Response to All Reviewers

We sincerely thank all reviewers for their thoughtful and constructive feedback. We would like to highlight our new cantilever evaluation (Figure 23), addressing concerns about more complex changes in solution multiplicity (CWjv, nVqx). This benchmark exhibits up to 20 branch-count transitions, and our method recovers the evolving solution sets at lower multiplicities while capturing the overall bifurcation structure in more challenging settings.

We are encouraged that the reviewers recognize the following strengths of our work:

  • Addressing an important limitation of single-output physical surrogates, enabling the learning of multiple admissible solutions from branch-unlabeled observations (CWjv, 1P5Z, nVqx).
  • Broad applicability across model families, demonstrated with FNOs, neural fields, and conditional diffusion (CWjv, 1P5Z, nVqx).
  • Meaningful improvements in multi-solution prediction, supported by experimental results across physical benchmarks (CWjv, 1P5Z, nVqx).
  • Complementary experimental analyses, including multimodal baselines, rare-branch imbalance, downstream inverse inference, and discussions of limitations (CWjv, 1P5Z).

In response to the reviewers' suggestions, we have further strengthened the manuscript through:

  • Additional baseline comparisons: rMCL, distance-based deduplication, and fixed-subset activation (CWjv, 1P5Z, nVqx).
  • More challenging solution multiplicities: A new cantilever benchmark and near-bifurcation analysis (CWjv, nVqx).
  • Robustness analyses: Clustering-threshold sensitivity and activation performance across solution multiplicities (1P5Z, nVqx).
  • Experimental consistency and reproducibility: Corrected diffusion ablations, clarified candidate budgets, and documented data splits and clustering thresholds (CWjv, 1P5Z, nVqx).
  • Improved clarity: Revised Figure 2 and expanded discussions of activation behavior and limitations (CWjv, nVqx).

We hope these additions and our detailed individual responses address the reviewers' concerns and further clarify our contributions. We sincerely appreciate their feedback and welcome additional suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.