acceptodds
Under review as a conference paper at ICLR 2027

Tangent Schrödinger Bridge Matching: Learning Stochastic Transport with Mechanistic Sensitivities

Abstract

Predicting how stochastic systems respond to changes in viscosity, reaction rates, or external forces requires costly simulations, motivating reusable learned models. Yet matching observed outcome distributions does not ensure accurate intervention responses. We introduce **Tangent Schrödinger Bridge Matching (Tangent-SBM)**, which learns stochastic transports from endpoint observations and mechanistic sensitivities. It propagates parameter derivatives alongside trajectories and supervises them against supplied targets. For average-response targets, single-rollout squared error also penalizes response variability; our objective uses two independent rollouts to match the mean without this additional penalty. We establish conditions under which sensitivity accuracy bounds finite-change prediction error and decision regret. Across Gaussian, stochastic double-well, PDEBench reaction–diffusion, and stochastic Navier–Stokes systems, Tangent-SBM improves sensitivity and finite-change prediction over matched conditional-bridge baselines while maintaining comparable endpoint and distributional accuracy. Controls examine target correctness, response objectives, and simulator-budget allocation. To test decision usefulness, we evaluate calibrated viscosity selection in Navier–Stokes: Tangent-SBM reduces tracking error relative to taking no action on every evaluated task.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Schrödinger bridges, diffusion models, stochastic differential equations, stochastic dynamics, mechanistic sensitivities, derivative supervision, scientific machine learning, intervention response

Reviews

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

This paper proposes Tangent Schrödinger Bridge Matching, which uses endpoint observations and mechanistic sensitivity information to jointly learn stochastic outcome distributions and their responses to parameter interventions. Within a conditional DSBM framework, the method propagates parameter derivatives along generated trajectories and distinguishes individual-trajectory response targets from mean-response targets according to the available supervision. The paper also establishes conditional relationships between sensitivity error, finite-intervention prediction error, and decision regret, and evaluates the method on four classes of systems and a calibrated viscosity-selection task.

Critical strengths
  1. The research question is valuable, and the motivation is clear.The paper explicitly distinguishes fitting observed outcomes from predicting the effects of interventions. For tasks that rely on models to adjust physical parameters, good endpoint predictions alone are insufficient; accurate parameter-response predictions are equally important. Using mechanistic sensitivities as additional supervision directly addresses this objective.

  2. The objective is specifically designed for mean-response targets.The authors explain why directly fitting each stochastic trajectory’s response to a mean target introduces an additional penalty on response variance, and use a two-rollout loss to remove this term. The design is supported by both a clear mathematical explanation and ablations comparing single- and two-rollout losses, providing useful insight into how stochastic models should incorporate mean-response supervision.

Critical weaknesses
  1. The distinct technical contribution beyond existing derivative supervision is not sufficiently compelling.Derivative supervision, stochastic sensitivity propagation, and independent double sampling all have established methodological foundations. Applying these ideas to response learning in stochastic bridges and distinguishing individual-trajectory from mean-response targets is a reasonable extension. However, the contribution currently appears more like an objective tailored to a particular framework than a sufficiently substantial standalone methodological advance.

  2. The theory primarily establishes conditional error relationships and provides limited support for the reliability of the practical learning procedure.Proposition 2 requires sufficiently accurate sensitivities along the entire intervention segment, rather than only at its starting point or selected training locations. Proposition 3 relies on a uniform upper bound on prediction errors across candidate actions. These results explain the method’s rationale, but do not sufficiently establish the practical reliability and generalization of the trained models.

Questions for authors
  1. With the same sensitivity targets and comparable training budgets, what advantages does Tangent-SBM offer over response-supervised direct predictors and other stochastic generative models that cannot be explained by derivative supervision alone? Can the authors add corresponding response supervision to conditional flow matching to better distinguish the contributions of the stochastic-bridge framework and the supervision objective?

  2. Can the authors evaluate sensitivity errors along the actual intervention segments and analyze the respective effects of target-approximation error, supervision coverage, and local linearization error on finite-intervention predictions and decision performance? How informative are the proposed error bounds quantitatively on the tested tasks?

  3. Can the authors include key baselines such as Sobolev-DSBM under the same response-supervision and scale-calibration protocol, and report both uncalibrated and calibrated decision results? Do the current advantages persist across more independent contexts, larger intervention ranges, or intervention-selection tasks involving multiple parameters?

5 comments

We thank the reviewer for recognizing the importance of intervention-response learning and our mean-response objective. We address all three questions with new experiments in Secs. 5.3–5.4 and Appendices H–J.

1. Contribution beyond derivative supervision.

Our contribution is target-aware response supervision for stochastic bridges, applied to free prediction-time rollouts (Secs. 3.3–3.4). We distinguish realization-level from mean-response targets: single-rollout supervision against a mean target penalizes response variability, whereas our two-rollout objective removes this penalty (Eqs. 8–10). The key methodological advance is supervising intervention responses in a way that respects the stochastic nature of the learned transport and the meaning of the supplied sensitivity targets.

With identical sensitivity targets, Tangent-SBM reduces SNS sensitivity error from 2.624 (Sobolev-DSBM) to 1.002 (Table 1; Appendix E, Table 9). On double well, mean-response RMSE decreases from 0.104 to 0.066, retaining greater response variability (Table 8). Thus, the gains extend beyond derivative labels alone.

We also added response-supervised Conditional Flow Matching (CFM), matching sensitivity targets, the two-rollout objective, and 8,000 optimizer updates. Supervision improves CFM's SNS ID error from 8.561 to 3.395, versus 1.002 for Tangent-SBM (Sec. 5.3; Appendix H.1, Table 17). This comparison tests the generative framework under shared response supervision, while Sobolev-DSBM isolates the supervision objective within the bridge framework. Point #3 below further compares response-supervised direct predictors on the same held-out decision task.

2. Intervention-segment accuracy and quantitative theory.

We empirically evaluated the assumptions underlying Propositions 2–3 (Sec. 4; Appendix J):

  • Segment accuracy: Proposition 2 requires accurate sensitivities along intervention paths, not only at their starting points. Across nine offsets in ([-0.25,0.25]), Tangent-SBM has lower error at every point: 57–65% below Conditional DSBM and 52–61% below TSBM (J.5, Table 27). Finite-change error also decreases to 1.073, versus 2.720/2.466 (Table 28).

  • Reference approximation: Increasing simulator reference pairs from 16 to 128 reduces relative sampling uncertainty from 14.7% to 8.6%. Tangent-SBM achieves (E_J=0.992), versus 2.696/2.489, retaining its advantage across three finite-difference steps (J.1, Tables 20–21).

  • Coverage: Increasing additional response-location coverage from 25% to 100%, while retaining endpoint-location labels, lowers segment sensitivity error from 1.326 to 1.058 and finite-change error from 1.349 to 1.073. Calibrated tracking remains 0.268–0.272 (J.6, Table 29).

  • Linearization and bounds: Since decisions approximate finite changes using local sensitivities, we measured the resulting approximation error: the physical linearization remainder averages 13.2% of the true finite change. Proposition 3 bounds how much worse the selected action can perform than the simulator oracle. The bound holds for every tested decision; mean normalized squared regret is 0.034, versus a mean bound of 0.207 (Appendix A.7; J.5, Table 28).

These results assess sensitivity accuracy along sampled intervention segments and its practical decision implications.

3. Decision controls and generality.

We expanded SNS evaluation to 64 held-out contexts (48 new), two goals each, and three seeds. All methods share actions, simulator references, and a calibration scale fitted on 16 separate development contexts and frozen for testing (Sec. 5.4; Appendix B.5, J.2).

MethodUncalibratedCalibrated
Sobolev-DSBM0.9330.296
Response-supervised direct predictor0.9950.562
Response-supervised CFM0.9310.636
Tangent-SBM0.4900.266

Tangent-SBM has the lowest tracking error before and after calibration, improving over no intervention on all 384 calibrated decisions (Table 3; J.2, Tables 22–23).

With calibration frozen, tracking error remains 0.250–0.263 across intervention magnitudes 0.0625–0.5, including unseen and doubled magnitudes. Tangent-SBM outperforms both baselines at every tested magnitude and improves over no intervention on all 480 decisions (J.4, Table 25).

We further added joint three-parameter PDEBench intervention selection over 27 actions. Relative to Conditional DSBM, Tangent-SBM reduces tracking error by 25% uncalibrated and 28% calibrated, and normalized regret by 79%, extending the findings beyond single-parameter decisions (J.4, Table 26).

Thank you for the new experiments. The matched CFM and Sobolev-DSBM comparisons clarify the contribution of the two-rollout objective, and the uncalibrated results strengthen the evidence for decision-making.

I am keeping my score for now. The refined Navier–Stokes sensitivity error is still close to the zero-response baseline (0.992 vs. 1), and the finite-change error in the segment study remains above it (1.073). The new analyses assess Propositions 2 and 3 on the tested cases, but reliability under finite, noisy supervision remains unclear.

  1. Could you vary the precision of the training sensitivity labels while holding the evaluation references fixed, and report the effects on finite-change error and decision regret?
  2. Could you report paired uncertainty for the calibrated tracking advantage over Sobolev-DSBM (0.266 vs. 0.296) and include that baseline in the larger-intervention tests?

Thank you for the constructive follow-up. We completed both requested evaluations to test robustness to noisy sensitivity supervision and the reliability of our decision improvements over Sobolev-DSBM.

1. Training-label precision with fixed evaluation references.

To test whether Tangent-SBM remains useful with noisy sensitivity labels, we varied the number of paired simulator estimates averaged per training label (1, 2, or 4 pairs), while keeping training locations, optimization settings, and evaluation references fixed. Each precision setting was retrained, with calibration fitted separately on the same development set and frozen before testing. Results are mean ± sample standard deviation across three seeds (Appendix J.7, Table 30).

Pairs per labelFinite-change errorUncalibrated regretCalibrated regret
11.138 ± 0.0950.346 ± 0.0600.082 ± 0.040
21.165 ± 0.0590.319 ± 0.0580.072 ± 0.012
41.052 ± 0.1490.251 ± 0.0720.058 ± 0.008

Increasing training-label precision from one to four pairs reduces mean regret by 27% uncalibrated and 29% calibrated. Finite-change error is lowest with four pairs, though not monotonic. Importantly, even with just one noisy pair per label, calibrated decisions improve over no intervention in 94/96 seed–task evaluations; with two or four pairs, they improve in all 96. Although finite-change error remains above the zero-response reference, these results demonstrate useful intervention selection under noisy supervision, with more precise labels further improving regret (Appendix J.7, Table 30).

2. Paired uncertainty and larger interventions versus Sobolev-DSBM.

To determine whether the calibrated tracking advantage (0.266 vs. 0.296; Sec. 5.4, Table 3; Appendix J.2, Table 23) is consistent across test contexts, we computed a 95% paired-bootstrap interval for Tangent-SBM minus Sobolev-DSBM. Across 2,000 context resamples, stratified by intervention value and keeping both goals and all three trained models paired within each context, the mean difference is −0.0304, with interval [−0.0427, −0.0189]. The entire interval is negative, showing that Tangent-SBM's lower tracking error persists under test-context resampling. This measures context uncertainty conditional on the trained models and fixed simulator estimates (Appendix J.2).

We also extended the larger-intervention experiment to Sobolev-DSBM, using identical simulator references and the original frozen calibration scales. Tangent-SBM achieves lower mean tracking error and regret at all five tested magnitudes. For the two largest:

Maximum magnitudeTracking: Tangent / SobolevRegret: Tangent / Sobolev
0.3750.250 / 0.2930.047 / 0.078
0.5000.260 / 0.3010.053 / 0.083

These represent 15% and 14% lower tracking error, and 40% and 36% lower regret, respectively, without recalibration. Thus, the decision advantage over the same-target Sobolev-DSBM control extends to larger interventions, rather than depending on the original action magnitudes (Appendix J.4, Table 25 and accompanying discussion).

Together, these experiments provide additional evidence that Tangent-SBM supports useful decisions under finite, noisy sensitivity supervision and that its calibrated advantage over Sobolev-DSBM is robust across the evaluated contexts and intervention magnitudes.

Reviewer_UJ5BOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes a method, Tangent Schrödinger Bridge Matching, that is designed to learn a stochastic transport to mimic the endpoints of a stochastic simulator, and to predict how this will change in response to interventions on the simulator parameters. They supervise the Schrödinger Bridge with simulation-parameter derivatives, to improve the estimated intervention responses. Across a few settings, they show that this method can better predict simulator responses to parameter changes than some baselines, and that this can lead to better decision-making on the ground-truth simulator.

Critical strengths
  • The problem, biasing learned stochastic transports to better capture interventions, is a compelling one.
  • The proposed two-rollout loss to based on parameter derivatives deals with the variance collapse of the single-rollout loss.
Critical weaknesses
  • The writing and presentation is generally confusing and difficult to follow. This largely stems from the minimal introduction and elaboration on the related DSBM work, which appears to be the foundation for this paper's method, with a new loss added. How DSBM learns a Schrödinger Bridge, though, is not well covered (e.g. what is Lbm\mathcal{L}_{bm} in Eq. 11?), which is should be given it's crucial relevance to the proposed method.
  • Figure 1 is very abstract, and does not do a good job of clearly showing the mechanics of the proposed method.
  • The work could be better placed amongst the broader related literature on biasing neural ODE/SDEs, or similar, based on inductive biases. In this case, the inductive bias is based on matching simulator parameter derivatives, to improve the modelling of interventions, but the work should more clearly acknowledge established ideas of biasing such models to improve inference. E.g. works that penalise trajectory models for traversing off-manifold areas [1,2], or for having instantaneous velocities that disagree with auxiliary data [3] seem related.
  • The matched simulator-budget experiments are important, as they can fairly elucidate the gain of the proposed matching loss. These should be in the main text, and done across all baselines.

[1] Kapuśniak, Kacper, et al. "Metric flow matching for smooth interpolations on the data manifold." Advances in Neural Information Processing Systems 37 (2024): 135011-135042.

[2] Amad, Harry, and Mihaela van der Schaar. "Hyperparameter Trajectory Inference with Conditional Lagrangian Optimal Transport." The Fourteenth International Conference on Learning Representations (2026).

[3] Petrović, Katarina, et al. "Curly flow matching for learning non-gradient field dynamics." Advances in Neural Information Processing Systems 38 (2026): 121544-121573.

Questions for authors
  • Can the authors provide an algorithm that clearly defines the proposed training method, and highlight in that what is different to typical DSBM training?
  • Can the authors more clearly place the method amongst the related literature by extending the Related Works section to better acknowledge existing works with similar high-level motivations.
  • Can the authors conduct the matched-budget experiments, from Appendix G, with all baseline models?
5 comments

We thank the reviewer for recognizing the compelling motivation and the value of our two-rollout objective. We have revised Secs. 2, 3.4, and 5.3, and Figure 1 to improve methodological clarity and positioning.

Q1 DSBM explanation, algorithm, and Figure 1. We have expanded Sec. 3.1 to explain conditional DSBM, while Sec. 3.4 now explicitly defines the DSBM bridge-matching loss used in Eq. 11 and provides Algorithm 1. Briefly, conditional DSBM iteratively learns a stochastic transport between endpoint distributions using drift-matching objectives, conditioned on context and intervention (u). Tangent-SBM adds tangent propagation and response supervision to standard conditional DSBM:

  1. Bridge matching: Train the conditional drift using the standard DSBM objective.
  2. Tangent propagation: At scheduled response updates, sample ((c,u,v)), generate free model rollouts, and propagate intervention sensitivities through the learned drift (Eqs. 3-4).
  3. Response supervision: Match terminal tangents to realization-level targets (Eq. 6), or use two independent rollouts to match conditional-mean targets (Eqs. 9-10), augmenting the drift-matching objective (Eq. 11).

Crucially, supervision is applied to free rollouts, ensuring that it constrains the intervention response of the actual prediction-time sampler rather than an endpoint-constrained training trajectory. We have also redesigned Figure 1 to illustrate this procedure and distinguish endpoint from response supervision.

Q2 Novelty and related work. We expanded Sec. 2 to explicitly situate Tangent-SBM within prior work incorporating inductive biases into learned neural dynamics. Our distinction is the quantity being supervised: rather than trajectory geometry or instantaneous velocities, Tangent-SBM constrains how the prediction-time sampler's outputs respond to parameter interventions, using simulator sensitivities. Metric Flow Matching [1] and Conditional Lagrangian Optimal Transport [2] introduce trajectory-level geometric constraints, while Curly Flow Matching [3] incorporates dynamical information into learned vector fields. These methods share our broader motivation of incorporating auxiliary structure beyond endpoint matching, but target different properties of the learned dynamics.. Our target-aware formulation further distinguishes realization-level from mean-response supervision; for the latter, the two-rollout objective avoids the response-variance penalty of single-rollout matching (Eqs. 8–10, Sec. 3.3). Appendix E, Table 8 demonstrates the benefit (RMSE 0.066 vs. 0.104), with greater retained response variability.

Q3 Expanded matched-budget experiments. We extended LocalBudget/SpreadBudget to Conditional Flow Matching, direct mean predictors, GSBM, and TSBM on SNS (Sec. 5.3, Table 2; Appendix G.1, Table 14). Each augmented baseline receives 2,592 additional simulator trajectories, matching response-label construction, with 8,000 optimizer updates fixed across allocations.

BaselineLocalBudget (E_J) SpreadBudget (E_J) 
Conditional Flow Matching5.5834.574
Direct mean predictor32.78627.696
GSBM3.1253.690
TSBM2.6992.693

Tangent-SBM achieves (E_J=1.002) under the matched ID evaluation. Thus, neither additional endpoint allocation closes its sensitivity advantage. These results complement the existing conditional-bridge controls on PDEBench and SNS (Appendix G, Table 13); on PDEBench, increasing SpreadBudget training updates also does not eliminate the gap. We additionally recorded training costs, including GSBM/TSBM path fitting. The key budget comparison appears in Sec. 5.3, Table 2, with full results and variability in Appendix G.1, Table 14, and compute accounting in Table 15.

I appreciate the effort gone into addressing the above points, and my concerns are mostly resolved. I still believe the writing quality can be improved, and particularly that the foundations for the method should be better explained, without delegating to references. For example, the endpoint coupling and drift-matching target from LBM\mathcal L_{BM} are mentioned briefly in the sentence after the loss is introduced, but not really explained at all. The explanation of how these are constructed, and how this leads to learning a Schrodinger bridge, are seemingly left to the original DSBM paper. Since this part of the learning process is still fundamental to the proposed method, I believe the original DSBM method should be elaborated on in more depth (possibly in an appendix). This would help the paper be a more complete work, rather than requiring readers to have read this prior work.

Thank you for clarifying this remaining concern. We have expanded Sec. 3.4 and added a self-contained appendix subsection (Appendix B.7), "DSBM background and bridge-matching construction," so readers can follow the procedure without consulting prior work.

The subsection explains how endpoint pairs are constructed and updated, gives the intermediate bridge-state sampling formula and explicit forward/backward drift targets, and connects these directly to qkq_k, VkV_k, and LBM\mathcal{L}_{\mathrm{BM}}. It then explains why the fitting steps are repeated and how the underlying iteration relates to the Schrödinger bridge.

We also clarify Tangent-SBM's central contribution: learning intervention responses alongside outcome distributions through tangent propagation and target-aware supervision of the prediction-time sampler.

Reviewer_LCS4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Tangent-SBM, which adds mechanistic sensitivity supervision to conditional Schrödinger bridge training to improve predicted responses of stochastic systems to parameter changes. It propagates directional parameter derivatives along sampling trajectories and uses a two-rollout objective for mean-response targets. Experiments cover four systems, plus a viscosity-selection study on stochastic Navier–Stokes (SNS).

Critical strengths
  1. The problem is clearly motivated: fitting outcome distributions does not guarantee correct parameter responses (Prop. 1, Gaussian experiment).
  2. Eqs. (8)–(10) explain the response-variance penalty that single-rollout squared loss introduces for mean-response targets, and how the two-rollout objective removes it. ID ablations support this: double-well mean-response RMSE drops from 0.104 to 0.066 (Table 6), and SNS sensitivity error is 1.002 versus 2.624 for the single-rollout variant (Table 7).
  3. The controls are thorough: shuffled targets, sign-reversed targets, rollout construction, and simulator-budget allocation.
Critical weaknesses
  1. Response prediction on SNS remains inaccurate. The ID sensitivity error is 1.002, close to the zero-response reference of 1. The finite-change error is 1.291, above that reference (Table 1, Sec. 5.2). Both worsen under OOD (Table 3), despite improved alignment and magnitude (Table 10).
  2. The added value of the bridge framework is insufficiently demonstrated. The main components build on prior work (Sec. 2). Direct response-supervised predictors perform better on some tasks (Appendix H). PDEBench is deterministic (Sec. 5.1), which leaves SNS as the only high-dimensional stochastic benchmark. The decision study uses only mean responses. Tangent-SBM does outperform the tested direct operators on SNS.
  3. The decision study does not adequately validate conditional response learning. It uses 16 test initial conditions, a single parameter, and candidate actions {0, ±0.125, ±0.25}. Targets and calibration share the same ±0.25 displacement (Appendix B.5), so generalization across magnitudes is untested. A baseline that keeps parameter dependence but ignores the initial state (e.g., regressing training response labels on u alone) is missing. Without it, gains from conditional response learning cannot be separated from shared response patterns and calibration.
Questions for authors
  1. Please report decision results without α. Please also report the sampling error of the SNS reference sensitivities (16 paired simulations), e.g., via bootstrap over existing pairs.
  2. How does the u-only baseline perform on sensitivity error and in the decision study?
  3. When only mean responses matter, what concrete advantage does Tangent-SBM offer over direct response-supervised predictors?
  4. Near-OOD lies within the sensitivity-supervision range for Gaussian, double well, and PDEBench (Appendix B). Could the main text distinguish “beyond endpoint supervision” from “beyond all supervision”?
  5. What is the additional training time and memory cost?
2 comments

We thank the reviewer for recognizing the motivation, theory, and ablations. New experiments in Secs. 5.1, 5.3–5.4 and Appendices B, H, J address the questions.

  1. SNS accuracy and reference uncertainty. SNS reference sensitivities use paired simulator runs with shared random forcing. To assess reference uncertainty, we increased pairs per condition from 16 to 128, reducing sampling uncertainty from 14.7% to 8.6%. With these refined references, Tangent-SBM achieves EJ=0.992E_J=0.992, versus 2.696 for Conditional DSBM and 2.489 for TSBM (63% and 60% reductions). Cosine alignment is also higher (0.609 vs. 0.136/0.060), indicating better directional agreement. We further resampled the 128 simulator pairs with replacement 2,000 times, holding models and evaluation conditions fixed. Both 95% bootstrap percentile ranges for error differences are entirely negative, confirming Tangent-SBM's advantage under reference resampling. The advantage also holds across three finite-difference step sizes (Appendix J.1, Tables 20–21).

  2. The requested uu-only baseline. To test whether Tangent-SBM learns responses specific to each initial condition rather than an average response for each intervention, we compare it with a regressor receiving only uu. This baseline achieves ID EJ=1.21E_J=1.21 and 61% oracle agreement, versus 1.002 and 75.8% for Tangent-SBM. We subtract the average response across initial conditions at each fixed uu, leaving context-specific deviations. A uu-only model predicts identical responses for all initial conditions and cannot capture these deviations (centered relative error 1.0). Tangent-SBM achieves 0.721 centered error and 0.532 cosine alignment, reducing squared error by 48.0%. All three seeds improve, showing the gains are not solely due to learning shared intervention-response patterns (Sec. 5.4; Appendix J.3, Table 24).

  3. Decision validation and calibration. We expanded testing from 16 to 64 held-out contexts, with calibration frozen from a separate development set. Tracking errors (lower is better) are:

MethodUncalibratedCalibrated
Conditional DSBM1.0000.499
TSBM0.9830.737
Tangent-SBM0.4900.266

Tangent-SBM improves over no intervention on all 384 calibrated seed–task decisions (Sec. 5.4, Table 3; Appendix J.2, Table 23). With calibration unchanged, we tested action magnitudes 0.0625–0.5, including unseen magnitudes and twice the calibration displacement. Tracking error remains 0.250–0.263, with improvement on all 480 seed–task decisions (Appendix J.4, Table 25).

  1. Advantages over direct predictors and other generative models. Using the same 64 test contexts, candidate actions, simulator references, and calibration protocol, Tangent-SBM achieves tracking error 0.266, versus 0.562 for the response-supervised direct predictor and 0.636 for response-supervised conditional flow matching (CFM)—reductions of 53% and 58% (Sec. 5.4, Table 3; Appendix J.2, Table 23). With identical sensitivity targets and the two-rollout mean-response objective, CFM achieves SNS ID sensitivity error 3.395, versus 1.002 for Tangent-SBM under the same evaluation protocol (Sec. 5.3; Appendix H.1, Table 17). Thus, Tangent-SBM improves response accuracy and intervention selection while retaining stochastic outcome generation. Appendix D reports comparable distributional accuracy to Conditional DSBM; Appendix E isolates the two-rollout benefit.

  2. Supervision coverage. We clarified the distinction between generalization beyond endpoint supervision and beyond all supervision in Sec. 5.1. For Gaussian, double well, and PDEBench, Near-OOD lies outside endpoint training but within sensitivity-supervision coverage. SNS Near-OOD (u=±1.25u=\pm1.25) lies outside both ranges ([−1,1][-1,1]), testing extrapolation beyond supervised intervention values (Appendix B.2).

  3. Computational cost. Response supervision requires differentiating through stochastic rollouts but occurs in only 160 of 8,000 training updates. On an RTX 6000 Ada (FP32, batch eight), per-seed median timings averaged across three seeds imply 4.9x estimated optimizer-loop cost relative to Conditional DSBM. Peak memory is 13.21 GiB during response updates versus 0.21 GiB otherwise. Ordinary outcome sampling remains unchanged (~18.2 ms/batch). Importantly, Sobolev-DSBM incurs comparable response-update time and memory, yet Tangent-SBM achieves 62% lower SNS sensitivity error and better calibrated tracking, demonstrating greater benefit at comparable response-update cost (Appendix B.6, Table 4; Tables 1 and 3).

Thank you again for your review. We wanted to briefly follow up on our response and manuscript revisions addressing SNS reference uncertainty, the intervention-only baseline, decision validation, additional comparisons demonstrating Tangent-SBM's advantages over direct predictors and response-supervised Conditional Flow Matching, and computational cost. If you have an opportunity to revisit these results, we would appreciate hearing whether they address your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.