acceptodds
Under review as a conference paper at ICLR 2027

Decision-Localized Characteristic Defects for Stochastic Neural Operator Selection

Abstract

Two stochastic neural operators can agree on global trajectory statistics yet predict different outcomes for the same path-dependent decision. With a small diagnostic-rollout budget, estimating these outcomes directly is noisy. Decision-localized characteristic defects (DLCD) select among pretrained operators by measuring exposure, drift, and covariance-sensitive increment errors in reference-defined decision regions. A fixed bank of physical observables, reusable reference calibration, and trusted local transitions provide the selection signal. At 512 diagnostic paths per candidate-condition, mean held-out regret across eight newly trained stochastic Allen–Cahn pools is 0.02510, compared with 0.02880 for regularized direct event estimation and 0.02693 for localized coefficient matching with matched information sources. The matched-information comparison yields 4.2% lower regret on a second Allen–Cahn cohort and 9.3% on six stochastic 2-D Kolmogorov pools. A history-conditioned decomposition explains the connection between local discrepancies and path decisions; ablations examine localization and aggregation. The gains are concentrated at smaller rollout budgets. Converting measured runtimes into additional direct-estimation paths produces an effective tie.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
stochastic neural operators, neural operator selection, stochastic differential equations, scientific machine learning, model selection, uncertainty quantification, physics-informed machine learning, operator learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1R78OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to select pre-trained stochastic neural operators for specific path-dependent events. DLCD scores candidate models using local drift, covariance-sensitive increment statistics, and occupancy in reference-defined regions. The experiments suggest that the method can improve model selection when only a limited number of diagnostic trajectories are available.

Critical strengths

Good agreement on global statistics does not necessarily mean that a model can accurately predict threshold-crossing events, so this is a useful problem to study. The comparison with direct event estimation is particularly helpful. The budget sweep shows when DLCD helps and where its advantage begins to weaken. The second Allen–Cahn cohort provides additional evidence. I also appreciate that the paper reports pools and conditions where DLCD performs worse.

Critical weaknesses

Appendix M states that the released records do not include the original evaluation horizon, the reference population used to estimate the quantiles, or the interpolation convention. These choices define the event being measured. Without this information, I find it difficult to assess the reported 9.3% improvement. The proposed replication protocol is useful, but it does not establish which settings were used in the reported experiments.

DLCD includes an occupancy component, while P2 does not. The paper acknowledges this and proposes adding that component to P2. I would like to see this experiment carried out, since it seems important for understanding how much of the improvement comes from the new residual construction.

Questions for authors

Table J.3 reports a regret of 0.0292 when 4,096 paths are allocated uniformly across eight candidates. This appears to correspond to 512 paths per candidate, but Table 1 reports 0.0251 at that budget. The two tables may use different evaluation protocols; if so, that difference should be clearly stated.

2 comments

Thank you for recognizing the importance of path-dependent selection, the direct-event estimator comparison, the budget sweep, the separately trained Allen-Cahn cohort, and the inclusion of adverse pools and conditions. We answer the three questions in the review directly.

1. Exposure-controlled P2: yes, DLCD retains lower mean regret in both cohorts. At frozen B = 512, the comparison on F-M is 0.02693 original P2 / 0.02610 P2 + exposure / 0.02510 DLCD; on independently trained N-U it is 0.03055 / 0.02990 / 0.02926. Relative to P2 + exposure, DLCD achieves 3.83% and 2.14% lower mean regret. The additional feature accounts for a meaningful part of the original P2 gap but does not eliminate it. The experiment addresses the reviewer's question using the already specified Appendix N comparator and without changing the submitted DLCD method. These are cohort means; original P2 favorable-pool counts are not relabeled as counts for the new comparator.

2. Kolmogorov event definition and interpretation. The original six KF pools test the enstrophy event in Section 5.3: cross a DEV-reference q90 threshold and persist above q85 for 0.05T. The reported six-pool regret is 0.03913 DLCD versus 0.04313 P2, with KF4 unfavorable. Appendix G specifies a distinct event-level replication protocol at T = 10 with a 201-point monitor, 0.5-unit dwell, and explicit interpolation/censoring. We do not equate these prospective choices with unidentified historical settings. The original released scientific-unit tables reproduce the reported six-pool comparison, while the paper's independent N-U cohort supplies a separate and fully specified held-out candidate-training replication. These historical and independent-replication protocols are distinct and serve complementary evidentiary roles.

3. Tables 1 and J.3: comparison of study definitions. The primary Table 1 result (0.02510) aggregates the fixed B = 512 per candidate-condition evaluation across C0-C3 and eight F-M pools. The Table J.3 uniform row (0.02920) belongs to a separately recorded fixed-total candidate-library allocation study with a 4,096-trajectory total budget. Therefore, 4,096/8 = 512 does not establish identical per-condition allocations, diagnostic draws, or pool-level selections. These two reported averages should not be treated as duplicate estimates from an identical experimental protocol. Table 1's primary result can be reconstructed from its pool-level records. The fixed-total release includes only policy-level aggregates, so it does not permit attribution of the 0.00410 difference to a specific execution factor. We will clarify the distinct study and budget definitions.

The exposure-controlled comparison strengthens the original fixed-budget evaluation, while the Appendix J allocation study remains a separate analysis rather than a replication of the primary fixed-B result. The new exposure-controlled experiment, the separately trained Allen cohort, and the original budget, localization, and adverse-case analyses jointly support the practical value of decision-localized physical selection in its stated regime.

Reviewer_1R78OpenReview ↗

Response to rebuttal

Thank you for providing the response. I decide to maintain my original score. While the authors provided additional ablation results, the method itself appears to have an overkilled complexity for the marginal performance gain with simple method like DIRECT-D1.

Reviewer_wU9fOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes decision-localized characteristic defects (DLCD) for selecting pretrained stochastic neural operators for path-dependent events. It combines occupancy, drift, and covariance-sensitive discrepancies in reference-defined regions with conventional trajectory diagnostics. Experiments report lower selection regret on Allen–Cahn and 2-D Kolmogorov problems at fixed rollout budgets, although the advantage over direct event estimation disappears in a timing-derived comparison.

Critical strengths

Relevant problem and plausible approach. Selecting a stochastic surrogate for a specific path-dependent decision is meaningful, and global trajectory statistics can miss errors relevant to that decision. Localizing physical increment diagnostics provides an interpretable approach to this problem.

Critical weaknesses

The central contribution is not convincingly established. The theoretical decomposition concerns true history-dependent continuation functions; it does not justify ranking models using the deployed signed, finite-witness score, whose residuals can cancel (Appendix H). This places substantial weight on empirical validation. However, the closest baseline, P2, omits DLCD’s exposure statistic, and P2 plus exposure is only proposed, never evaluated—leaving the source of the reported improvement unresolved. Furthermore, direct estimation effectively ties DLCD after the timing-derived budget adjustment: 0.02508 versus 0.02520 regret. A directly measured equal-time comparison remains missing (§5.2; Table 2; Appendix N).

Questions for authors

I think the paper can be enhanced greatly if the writing is improved. I think the introduction and some preliminary sections are either poorly written or simply omitted, and the writing style is not very "standard" which makes it difficult to read.

1 comment

Thank you for the careful review. The key question is whether the deployed finite-witness score enables decision-specific selection, rather than whether the continuation identity guarantees ranking. The submitted experiments and new exposure control address this directly.

1. The contribution is a post-training decision-specific selection criterion, not a new architecture or unrestricted error bound. DLCD selects among already-trained stochastic operators for a specified path-dependent outcome, using limited diagnostic rollouts without accessing held-out candidate decision losses. It combines (i) reference-defined decision-region exposure, (ii) calibrated local drift and covariance-sensitive characteristic discrepancies from fixed physical witnesses, and (iii) frozen NULL-REF calibration and selectively queried trusted one-step targets. It differs from global fidelity scores, P2's coefficient matching, and DIRECT-D1's direct estimation of event outcomes. These information roles remain separated (Sections 3-4; Figure 1). The submitted Sections 3-5 evaluate the original criterion against held-out decision losses, localization controls, direct estimators, independently trained candidate pools, and adverse cases. The additional experiment addresses the exposure-related attribution question without changing the proposed method.

2. The original submission links the practical diagnostic to the actual decision loss. Among 40 development C0 candidates, the characteristic score's Spearman association with true held-out D1 loss is 0.708, versus 0.476 for BASE; controlling for BASE rank and candidate family retains 0.531 (Section 5.1; Appendix D). Global characteristic audit A0 gives 0.02795 regret, versus 0.02510 for localized DLCD (Section 5.2). On frozen B = 512 held-out pools, regret is 0.02510 for DLCD, 0.03250 for BASE, and 0.02880 for regularized DIRECT-D1. N-U pools and adverse cases bound generality.

3. The exposure-feature alternative has now been tested directly. P2 + exposure, specified in the submitted Appendix N with frozen calibration, lowers original P2 mean regret from 0.02693 to 0.02610 on F-M and from 0.03055 to 0.02990 on independently trained N-U. DLCD remains at 0.02510 and 0.02926, respectively: 3.83% and 2.14% lower mean regret than strengthened P2. Exposure accounts for part, but not all, of the mean gap in either cohort. Cohort means alone establish neither paired significance nor individual drift/covariance attribution. The result is a focused attribution control, not a new DLCD method.

4. The theorem and finite-witness selector are distinct; cancellation is already tested. Proposition 1 is an exact history-conditioned continuation identity for a bounded path functional. Appendix H specifies witness-approximation and localization terms. The implemented signed finite-witness score is not a proven regret bound or a universal ranking certificate. The submitted Section 5.4 directly probes cancellation: regret is 0.02510 for signed DLCD, 0.02542 for a cellwise non-cancelling version (7/8 identical selections), 0.02618 for history-absolute aggregation, and 0.02705 for tail-max aggregation. The continuation-value fit gives R² = 0.58 overall and 0.66 inside A1. These empirically check the implemented criterion without excluding all cancellation.

5. The computational claim has a meaningful but explicit boundary. At matched B = 512 paths per candidate-condition, DLCD beats raw and regularized DIRECT-D1 on regret. Median online times are 104.7 s for DLCD and 44.1 s for DIRECT-D1. The paper also reports a timing-derived, not separately executed end-to-end, DIRECT-equivalent B = 1216 sensitivity: DIRECT-D1 regret 0.02508 versus the corresponding DLCD timing aggregate 0.02520, effectively tied. Separately, a measured equal-clock physical-baseline comparison yields P2 at B = 633: 104.6 s and 0.02641 regret; DLCD at B = 512: 104.7 s and 0.02520 (Section 5.2). These comparisons are distinct: no universal wall-clock superiority is claimed.

6. Presentation concerns can be resolved by foregrounding the original method. Section 1 states the basin-commitment decision; Section 3 gives a worked eight-candidate instance; Figure 1 shows the strict selection/evaluation information boundary; Section 3.3 and Appendix H distinguish theoretical motivation from empirical finite-witness ranking. The exposition should foreground this existing chain and its limits. The central claim is a decision-specific physical audit that improves observed held-out selection regret at restricted rollout budgets, with an explicitly reported compute boundary.

We respectfully ask that the submitted decision-loss evidence, localization and aggregation controls, independent training cohorts, adverse cases, and the new exposure-controlled comparison be assessed together.

Reviewer_RmowOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • This paper proposes Decision-Localized Characteristic Defects (DLCD), a criterion for selecting a pretrained stochastic simulator according to its accuracy on a downstream event, such as a transition between states. DLCD measures discrepancies in exposure, drift, and covariance near a task-relevant region, normalizes them using reference data, and combines them with conventional fidelity diagnostics.
  • The paper provides theoretical motivation through finite-step identities and evaluates simulator-selection regret on stochastic Allen–Cahn and Kolmogorov-flow problems, including a second Allen-Cahn cohort. The supplement provides reference implementations, processed results, and verification scripts. The reported results support the usefulness of task-specific diagnostics, although the contribution of the proposed characteristic defects is not fully isolated by the comparisons.
Critical strengths
  • A meaningful model-selection problem with a physically motivated approach. Selecting a stochastic simulator for a specific event under limited evaluation budgets is practically relevant. Exposure, drift, and covariance have a coherent connection to stochastic dynamics, and localizing their evaluation near task-relevant regions is a plausible way to detect errors that global fidelity metrics overlook.

  • Useful empirical replication and transparent supplementary material. The Allen-Cahn evaluation includes a second independently trained cohort, with trained model pools serving as the units of comparison. The reported improvement over P2 appears in both cohorts. The supplement supplies inspectable scoring and event-reconstruction implementations and processed results whose principal summaries can be reproduced. The paper also reports computational costs and settings where the advantage is limited, which helps assess its practical scope.

Critical weaknesses
  • The strongest comparison does not isolate the proposed contribution. DLCD includes an exposure discrepancy that the P2 comparator lacks. Consequently, the reported improvement does not establish how much comes from the proposed drift/covariance diagnostics versus the additional exposure information. The supplement provides an executable P2-plus-exposure comparator but explicitly states that it has not been evaluated. This comparison is consequential because the theoretical analysis motivates the construction without establishing that the resulting score reliably ranks simulators by downstream event error.

  • The historical Kolmogorov results remain insufficiently reproducible. The supplement explicitly acknowledges that the original evaluation horizon, the population used to estimate event-defining reference quantiles, and complete interpolation/censoring settings cannot be recovered from the released material. A prospective follow-up protocol is supplied, but it does not establish which settings generated the reported results. These choices define the target event and therefore affect the interpretation of the claimed transfer beyond Allen-Cahn.

Questions for authors
  • Can you evaluate the supplied P2-plus-exposure comparator on both Allen-Cahn cohorts using the same frozen calibration, diagnostic access, and evaluation budgets? Does DLCD retain its advantage?

  • Can you recover and document the exact horizon, quantile-estimation population, and event interpolation/censoring settings used for the original Kolmogorov results, with records linking those settings to the reported experiments?

  • The uniform allocation of 4,096 trajectories across eight candidates corresponds to 512 per candidate, but its reported regret is 0.0292, compared with 0.0251 in the main 512-per-candidate experiment. What differences in protocol or aggregation explain this discrepancy?

2 comments

Thank you for the constructive assessment and for recognizing the decision-specific selection problem, the physical motivation of DLCD, and the value of independently trained candidate pools and inspectable supplementary analyses. We answer the three evidentiary questions, distinguishing original findings from the new exposure-controlled experiment.

1. Does DLCD still improve on P2 after giving P2 the exposure feature? Yes, in both independently trained cohort means. The original comparison matched the candidate diagnostic trajectories, reference-defined region, trusted-dynamics information sources, and held-out selection boundary, but P2 did not contain DLCD's standalone exposure statistic. We evaluated the P2 + exposure score specified in the submitted Appendix N, with the reference calibration and scoring rule frozen rather than retuned using held-out outcomes.

B = 512, mean held-out regretF-M (8 pools)N-U (8 pools)
Original P20.026930.03055
P2 + exposure0.026100.02990
DLCD0.025100.02926
Further DLCD reduction vs. P2 + exposure3.83%2.14%

Adding exposure closes approximately 45% of the original P2-DLCD mean gap in F-M and 50% in N-U, so it is an important baseline improvement. Crucially, it does not eliminate the gap in either independently trained cohort. This supports an incremental cohort-mean selection advantage beyond adding exposure to coefficient matching; it does not isolate the effect of each constituent. The original 7/8 and 6/8 favorable-pool counts describe original P2, not the strengthened comparator; we do not conflate their statistics.

The original submission also provides complementary attribution controls: global A0 characteristic auditing yields 0.02795 regret compared with 0.02510 for reference-defined localized DLCD (Section 5.2), and localized non-characteristic controls perform worse. Together, these experiments provide complementary evidence from exposure-controlled attribution, decision localization, decision-loss association, independently trained cohorts, and downstream selection regret.

2. What is supported by the Kolmogorov experiment? Section 5.3 defines the enstrophy decision by an upward crossing of the DEV-reference 90th-percentile threshold followed by persistence above the 85th-percentile threshold for 0.05T. Table 3 reports all six KF pools, including adverse pool KF4: DLCD achieves 0.03913 mean regret versus 0.04313 for P2 (9.3% lower). The released six-pool data reproduce those reported aggregate and pool-level comparisons. The submission also specifies in Appendix G an independent, fully defined event-monitoring replication protocol (T = 10, 201 monitor observations, a 0.5-unit dwell, and explicit event interpolation and censoring). The original six-pool result and this independently specified replication protocol serve different purposes and should not be conflated. Independently of the cross-system result, the second Allen-Cahn cohort supplies a distinct candidate-training replication with a fixed, fully specified D1 event.

3. How do Table 1 and Table J.3 relate? Table 1 reports the primary fixed B = 512 per candidate per physical condition evaluation across C0-C3, yielding 0.02510 mean regret from eight F-M pools. With eight candidates, this corresponds to 16,384 candidate-condition diagnostic trajectories per pool. Table J.3 instead reports the uniform arm of a separate fixed-total candidate-library allocation study, with 4,096 total trajectories and aggregate regret 0.02920. Although 4,096/8 = 512, this quotient gives a nominal per-candidate library allocation, not evidence that every candidate-condition received B = 512 under Table 1's evaluation protocol. The two rows therefore should not be treated as identical-protocol replications. The primary Table 1 result is reconstructible from its pool-level records. The available fixed-total aggregate does not establish the precise cause of the 0.00410 difference, so we do not attribute it to specific draws or aggregation decisions without supporting records. We will distinguish the two budget definitions explicitly.

Finally, the theory concerns raised alongside attribution do not rest on a universal rank guarantee: Proposition 1 is an exact history-conditioned identity; the deployed finite-witness score is assessed empirically. Section 5.4 already reports regret of 0.02542 for a cellwise non-cancelling variant versus 0.02510 for signed DLCD, with 7/8 selection agreement, directly probing the cancellation issue. Taken together, the new controlled comparator and the original localization, replication, and aggregation experiments reinforce the contribution of the submitted criterion under its stated finite-rollout objective.

Thank you for the additional comparison and budget clarification. I appreciate the stronger evidence provided in the rebuttal. Given the modest gains over P2 + exposure, uncertainty about their consistency across pools, and the effective tie with direct event estimation in the timing-derived comparison, I’m keeping my rating at 2 (weak rejection).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.