acceptodds
Under review as a conference paper at ICLR 2027

Task-Locked Predictive Adaptation: Leftover Residual Densities from Few Target Labels

Abstract

A trained neural operator is often reused when the unresolved forcing of a PDE changes. Users already rely on a predictive distribution for one operational functional, such as a regional mean. Recalibrating the residual as an unrestricted joint density moves that functional. Few labeled target fields are available. The leftover residual still requires a density. We keep one shared density for the task coordinate. The labels update only the density of the remaining coefficients given the task. From the source we build a dictionary of these leftover models: scale and tail changes, coupling to the task, and both together. Sequential likelihood updates the weights. The deployed leftover density averages the sequential predictors. We call this task-locked predictive adaptation. The adapted predictor has the same task law as the shared task density. Under i.i.d. target sampling, with the dictionary fixed from the source, the expected excess log loss relative to the source leftover density is at most . Our analysis separates what the labels can buy from what the dictionary itself can represent. We test viscous Burgers with a frozen Fourier neural operator, ten training seeds, and 32 target labels. Residuals are scored on eight Fourier coefficients. Under a joint change of forcing tail and dependence, the leftover log score improves by 0.450 nats. Under no shift the mean change is 0.0005 nats. A maximum-likelihood fit of the same dictionary satisfies its calibration criterion in every recorded fit but scores 0.0717 nats higher on held-out fields. Composed corrections improve that score by 0.050 nats over a mixture of the separate families.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
uncertainty quantification, neural operators, distribution shift, scientific machine learning, predictive density estimation, mixture aggregation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gWA6OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper considers adapting a residual predictive distribution while keeping a supplied task marginal unchanged. Source information defines a family of conditional corrections, and a small target sample updates their weights. The analysis separates restrictions of the candidate family from estimation error, including a symmetry-based approximation barrier. Controlled PDE experiments examine the benefits and costs of the resulting adaptation procedure.

Critical strengths
  1. The theoretical separation is the strongest part of the paper. Exact preservation of a task law, the approximation error imposed by a shared symmetry, and an expected estimation guarantee address different questions. The paper generally keeps those questions distinct. In particular, it does not replace the individual-candidate oracle with a stronger convex-mixture guarantee that the estimator cannot satisfy.

  2. The experiments are reported with appropriate restraint. The authors show where prefix averaging loses to the final posterior, where broader directional coverage makes the aggregate result worse, and where improved log score does not improve an event prediction. Those outcomes help explain the construction's limits. Treating trained seeds as the unit of inference also avoids overstating the number of independent model replications.

Critical weaknesses
  1. The practical payoff is less established than the mathematical result. The main application is restricted to one PDE and a projected residual representation. The preserved task distribution is supplied to the conditional procedure, so exact preservation does not establish that this law is well calibrated. Likewise, the main paired log-score comparison removes that common marginal. The event results in Section 5.5 are mixed and provide only limited evidence for better operational decisions. I regard this as a limitation of application scope, rather than an error in the preservation claim.

  2. The estimator comparisons do not fully separate aggregation from source protection. The proposed construction gives substantial protection to the source reference, whereas the unregularized mixture fit and continuous alternatives do not match that protection. The advantage over those baselines is therefore harder to attribute to the chosen estimation rule. Moreover, the final posterior performs better than prefix averaging in the principal setting at several budgets. The authors acknowledge this, but it means the case for prefix averaging rests chiefly on its expected guarantee. A comparison with source-regularized fitting would make the practical choice easier to assess.

Questions for authors
  1. Can the absolute adaptation gains for the additional architecture be reported? The current difference-of-gains comparison does not by itself establish its effectiveness there.
  2. Was a source-regularized fit over the same candidate family evaluated? Existing results would help distinguish the contribution of source protection from the estimation rule.
  3. Is any assessment of the supplied task marginal's calibration or the absolute joint predictive score available? This would clarify what the conditional improvement means for the whole predictive distribution.
  4. How sensitive are the main gains to the prior mass assigned to the source reference? Please summarize any existing sensitivity analysis.
  5. What do the independent source-direction checks establish about selection stability? Please distinguish finite-difference accuracy from the nonlinear-response approximation discussed in Appendix A.16.
2 comments
AuthorsOpenReview ↗

Response to Reviewer gWA6

Your review asks what the conditional construction contributes in practice, and how much of the estimator comparison comes from source protection. We now follow those questions through to physical scores, absolute joint predictions and matched regularized fitting. The applications retain useful task forecasts and adapt the joint information those forecasts leave open. Matched comparisons then measure how fitting rules use the same family and source prior.

W1. Practical payoff and application scope

The revision evaluates complete Burgers regional predictions and adds weather, hydrology, advection–diffusion and particles (Section 5.4; Table 23). On G5, original C774 lowers tail OR-event Brier by 3.31×10−43.31\times10^{-4}, with improvement interval [1.31,5.30]×10−4[1.31,5.30]\times10^{-4}. Weather J121 prefix lowers gust CRPS by 10.7% and 3.49% at the two stations; the combined gain is 0.111 m/s [0.0127, 0.205]. The protocol's primary endpoint has gain 0.115 m/s with its own interval [-0.0126, 0.230]. In hydrology, retaining supplied volume model m1m_1 lowers volume and peak-flow CRPS by 26.3% and 41.7% relative to an additional 32-label update with shape fixed. With volume fixed, J13 endpoint and MAP4 improve peak-date RPS by 6.16% and 8.14%; HEFS shape remains strongest (Tables 24, 27 and 28). Table 3 also evaluates a frozen whole-joint branch-selection policy directly in event Brier risk.

W2/Q2. Aggregation versus source protection

We added MAP fits with strengths 4, 16 and 64 over the same C774 candidates, prior, labels, marginal and completion. Their objective is conditional log-likelihood minus τKL(π∥w)\tau\mathrm{KL}(\pi\Vert w). Realized reference weights make the comparison concrete: prefix, final and endpoint average 0.773, 0.704 and 0.707; MAP4, MAP16 and MAP64 average 0.665, 0.774 and 0.797 over all 448 G5 contexts. At almost equal mean reference weight, prefix has mean NLL 9.805 versus MAP16's 9.814 (Tables 2, 11 and 22).

The stronger tail/combined performance of other estimators remains explicit. On combined shift, NLL differences from source are -0.215 for prefix, -0.219 for MAP4, -0.225 for final and -0.227 for endpoint. Prefix retains the expected source-reference guarantee and has lower null cost than MAP4 (+0.00184 versus +0.0204). These matched comparisons explain the practical tradeoff behind the estimator choice.

Q1. Gains within the additional architecture

C774 improves over the custom convolutional backbone's own source reference in each of its five trained instances for all three main shifts. Mean NLL differences are -0.254 [-0.423, -0.0845] for tail, -0.0207 [-0.0273, -0.0141] for dependence and -0.323 [-0.501, -0.145] for combined. Each checkpoint uses its own source chart and response direction. Table 17 retains all six controls and the separate FNO difference-of-gains comparison.

Q3. Quality of the task marginal and absolute joint prediction

We report weather task, conditional and joint NLL together with marginal calibration. At Klippeneck, with pooled source HH fixed, an independent 2017 temperature update lowers task NLL from 1.828 to 1.793 and joint NLL from 4.517 to 4.482. Nominal 90% coverage moves from 95.5% to 93.2%. Conversely, with the supplied marginal fixed, J121-prefix adaptation lowers joint NLL from 4.525 to 4.253 (Table 25). For hydrological volume model m1m_1, nominal 50% physical equal-tail coverage is 0.938, and the corresponding central randomized-PIT frequency is 0.758; the latter accounts for the zero-flow atom (Table 29). These measurements separate task-forecast quality from conditional improvement.

Q4. Reference-prior sensitivity

Raising reference mass from 0.5 to 0.95 lowers null cost from 0.00681 to 5.10×10−55.10\times10^{-5} nats, while tail, dependence and combined gains remain 0.296, 0.0125 and 0.334 nats. Coarsening the radial grid from 17 to 9 nodes changes mean prefix NLL by at most 2.01×10−42.01\times10^{-4} nats. Tables 18–20 report all priors, grids and estimators.

Q5. What the independent direction checks establish

The local Jacobian rule has finite-difference relative error [3.13,3.39]×10−7[3.13,3.39]\times10^{-7} and reproduces the selected forcing pair in 20/20 independent source checks, with direction cosines 0.993–0.99993. Agreement with the full nonlinear covariance-response criterion is 16/20; at finite correlations ±0.2, the 40 checks have response discrepancies of 1.37–7.11%. Appendix A.9 uses Price's identity to explain the curvature terms separating those criteria. The additional covariance-response subspace is a separate representation extension: its cosines are 0.372, 0.656, 0.408 and 0.922, with three of four sources failing the combined angle/eigengap check (Table 21).

Reviewer_gWA6OpenReview ↗

Official Comment by Reviewer gWA6

Thanks to the author for the rebuttal. After reading the rebuttal carefully, I will maintain the score.

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

This research addresses the problem of unresolved forcing in partial differential equations (PDEs) by introducing a novel framework for model adaptation. The primary objective is to determine how much information regarding specific task characteristics, such as scale and tail behavior is derived from the underlying data representation versus the provided labels during the learning process. To address this, the authors propose "task-locked predictive adaptation." This method utilizes a dictionary of "leftover" models that capture variations in scales and tails across different tasks. The model weights are updated via a sequential likelihood approach, allowing the system to isolate the influence of specific features. The methodology was evaluated using a Fourier neural network applied to 32 target labels within the context of viscous Burgers' equations.

A central challenge identified in the study is determining the specific conditions under which model changes should be permitted during adaptation. To resolve this, the authors implemented a selection process where a subset of labels defines the candidate corrections prior to the adaptation phase. These contributions are structured as a conditional adaptation construction designed to preserve the underlying physical laws. The findings suggest that task-locked predictive adaptation successfully isolates the contributions of different data components, providing a robust mechanism for adapting models to new tasks while maintaining physical consistency.

Critical strengths
  • Robust Experimental Validation: The study includes a comprehensive set of protocols and validations that effectively demonstrate the practical impact of the proposed method on Fourier neural operations.

  • Rigorous Mathematical Framework: Equation 6 provides a well-defined error analysis, where the inclusion of conditional approximations serves as an essential safeguard for representation accuracy.

Critical weaknesses
  • Ambiguity in Loss Function Application: It is unclear whether the defined loss function is applied exclusively to the model weights or if it also governs the generation of the dictionary.

  • Lack of Justification for Parameter Selection: The criteria used to define the directions for the C774 construction are not specified, making it difficult to assess the methodology's reproducibility.

  • Insufficient Comparative Analysis: The evaluation lacks a comparison against relevant baseline methods, such as Meta-Learning or Bayesian Neural Networks, which would provide a more comprehensive context for the results.

Questions for authors
  • Please explicitly state the criteria and methodology used to construct the C774 directions.
  • Please include a comparison against standard baselines, such as Meta-Learning and Bayesian Neural Networks, in the evaluation section.
1 comment
AuthorsOpenReview ↗

Response to Reviewer BDJD

Your questions identified two places where the method needed a clearer description: what source data determine before adaptation, and what the target loss then learns. We have made that sequence explicit and added a conditional Bayesian neural baseline under the same task interface and label budget. The comparison now shows where the fixed dictionary is useful, while the construction details make the experiment reproducible.

W1. Which objects does the loss determine?

Source data determine the neural operator, residual coordinates, reference density, response direction, candidate transforms and prior. Once these objects are fixed, target conditional likelihood updates the mixture weights. For the original C774 estimator,

wt,j∝πj∏i=1thj(ui∣zi),wˉK=1K+1∑t=0Kwt.w_{t,j}\propto\pi_j\prod_{i=1}^{t}h_j(u_i\mid z_i),\qquad \bar w_K=\frac{1}{K+1}\sum_{t=0}^{K}w_t.

At 32 labels, deployment therefore averages 33 indexed predictors. The source reference is fitted by source likelihood, and its direction comes from source response calculations; neither is selected by the target adaptation loss. Continuous-head and BNN comparators learn their own conditional parameters at the same target budget (Sections 3.1–3.3). This distinction separates the construction of the available predictions from the target sample's choice among them.

W2/Q1. How is the C774 direction selected, and why this criterion?

The criterion measures which forcing pair produces the strongest task–residual cross-covariance response in the source linearization. On 128 source design inputs, central differences with step 0.025 estimate

cij=1128∑n=1128[Ju,i(n)JF,j(n)+Ju,j(n)JF,i(n)].c_{ij}=\frac{1}{128}\sum_{n=1}^{128}\left[J_{u,i}^{(n)}J_{F,j}^{(n)}+J_{u,j}^{(n)}J_{F,i}^{(n)}\right].

We select the largest ∥cij∥\|c_{ij}\| among all 28 forcing pairs, break ties lexicographically, and set v=cij/∥cij∥v=c_{ij}/\|c_{ij}\|. A correlation change ρ\rho has response ρcij\rho c_{ij} in that averaged linearization, which motivates the choice. Independent source checks reproduce the selected pair in 20/20 cases. The full nonlinear covariance-response criterion agrees in 16/20; Appendix A.9 explains the additional curvature terms (Section 3.1, Equation 3; Table 21).

Appendix B.4 also gives the whole candidate grid: 1 source reference, 85 radial alternatives, 8 directed transforms and 680 compositions. The radial alternatives use 17 scales from 0.25 to 4 and five degrees of freedom; the four shifts are ±0.4 and ±0.8, each with unit or compensating contraction. The four group prior masses are 0.8/0.05/0.05/0.10.

W3/Q2. Comparison with Bayesian neural and meta-learning baselines

We added a conditional Bayes-by-Backprop baseline using the same 2,048 source observations, 32 target labels, task marginal, completion and test fields. C774 prefix has lower mean NLL than the selected full-output BNN by 0.0968 nats, with descriptive four-source interval [0.0879, 0.106] for that improvement. The comparison favors C774 in all seven controls and all four source means (Tables 2 and 12). Below, negative differences favor C774:

G5 controlC774 minus BNN NLL
Null-0.0509
Tail-0.228
Positive correlation-0.0412
Negative correlation-0.0404
Combined-0.235
Weak-0.0508
Outside-0.0313

The BNN places mean-field Gaussian variational posteriors on all weights and biases of a 16-unit tanh network. Full and diagonal describe its output likelihood covariance; the full model has 627 mean parameters. Separate F2 development data selected the full likelihood and source-posterior prior-SD multiplier 3 before scoring the already revealed G5 matrix, so this is a matched supplemental comparison. Banks are averaged within each trained source before forming the reported intervals.

We also checked whether the result depends on predictive sampling or training budget. The main run uses 2,000 source and 600 target ELBO steps and 512 predictive weight draws. On the designated bank-0 subset, a second seed, 8,192 draws and doubled training preserve the comparison direction. Doubling training narrows C774 minus full-BNN NLL from -0.0811 to -0.0622 nats (Table 5). The diagonal-output BNN gives a similar main-matrix gap of -0.0937 nats. A cross-task MAML baseline remains unimplemented. Tables 2 and 11 place these Bayesian results alongside same-family source-regularized MAP and continuous Gaussian/Student heads, making the alternatives at the stated budget directly comparable.

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

This paper looks at how to adjust the joint distribution of residuals with just a few target labels after transferring a neural operator, while strictly keeping a pre-specified task marginal distribution intact. The authors build candidate conditional densities from source data and then update their mixture weights, analyzing the symmetry limits of the candidate dictionary. Burgers equation experiments show the method can adapt when both tails and dependencies change. But if the specified marginal is off to begin with, that error gets locked in, and the fixed dictionary can't capture every kind of target shift.

Critical strengths

1.Strictly preserves the specified task's marginal distribution when updating residual structure.

2.Candidate densities are pre-built from source data; only weights need updating with a few target labels.

3.Separates dictionary expressiveness limits from finite-sample estimation errors.

4.Mathematically characterizes the irreducible error caused by shared symmetries.

5.Honestly reports cases where prefix averaging and certain event predictions aren't dominant.

Critical weaknesses

1.If the specified task marginal is wrong, the method locks in that error.

2.The uniform conditional density doesn't leverage remaining information in the input x.

3.Fixed directions and symmetries can't capture certain types of target shifts.

4.The theoretical guarantee isn't relative to the best possible mixture density.

5.Experiments are mostly limited to controlled Burgers shifts and an 8D residual.

Questions for authors

1.When the target shift is unknown, how do you know the source-selected direction is still valid?

2.How does a mis-specified upstream task marginal affect joint prediction?

3.Can improvements in the 8D coefficients carry over to the full physical field?

4.How sensitive is the method to prior weight and candidate grid choices?

5.Which joint events benefit, and can you predict that in advance?

3 comments
AuthorsOpenReview ↗

Response to Reviewer AGNZ (1/2): Summary and Main Concerns

Your comments connect the task forecast we retain with the joint prediction we adapt. We have organized the revision around that connection. A supplied task forecast can be inspected and updated upstream; with that choice fixed, a small target sample adapts the remaining conditional structure. The candidate family determines which changes this adaptation can represent. The revised analysis quantifies that dependence, and the experiments follow it through to physical predictions. We address your five weaknesses below and your five questions in the accompanying comment.

W1. What if the supplied task marginal is wrong?

We now evaluate the task forecast as an auditable upstream module. This makes retaining it a choice whose value can be measured. In hydrology, keeping the supplied volume model m1m_1 lowers volume CRPS by 26.3% and peak-flow CRPS by 41.7% relative to an additional regularized 32-label update, with the conditional shape model and zero-flow probability fixed. Across the 128 weeks, the difference traces to a sparse-update episode whose increased variance amplifies the physical mean after the lognormal transformation. We also report weather marginal calibration and absolute joint scores. Section 4.1 exposes task misspecification as an additive risk term, so the workflow is to assess the task forecast upstream and then adapt the conditional law (Section 5.4; Tables 25 and 27).

W2. Can the conditional model use the remaining information in xx?

Yes. Equation 2 permits HD(u∣z,x)H_D(u\mid z,x) while preserving the task marginal at each input. The pooled C774 head uses HD(u∣z)H_D(u\mid z) to keep the fitted problem small. We now measure what that pooling costs with a matched input-dependent head: the same 32 labels, task marginal, coordinates and Student shrinkage strength. At Klippeneck, adding six current-input coefficients lowers gust CRPS from 1.588 to 1.453 m/s and joint NLL by 0.122 nats. The CRPS improvement interval is [0.0758, 0.190] m/s. Shrinkage strength 64 was selected by 2017 event Brier; this input comparison is a post-hoc diagnostic on 2018 data. At Würzburg, both the same-strength and separately selected comparisons have intervals crossing zero (Table 26). Section 4.1 identifies the associated pooling term as I(U;X∣Z)I(U;X\mid Z).

W3. What do fixed directions leave unrepresented?

We make this a quantitative design question. All C774 candidates share the orthogonal symmetries fixing the source direction vv, so every mixture inherits them. Proposition 2 separates the corresponding symmetry floor from the remaining invariant-class approximation and fitting error. In the seven-dimensional Gaussian example, adding the omitted variance direction removes a 0.0852-nat symmetry floor. Within the enlarged subspace, a matched covariance extension has KL 0.1058 at 32 labels and 0.02149 at 256 labels, versus 0.1121 and 0.1083 without covariance parameters. This connects representation capacity to the budget needed to use it. Separate PDE grid and 32-label deployment comparisons appear in Section 5.3 and Tables 7 and 16.

W4. Why is the guarantee not against the best mixture?

The original prefix estimator serves the few-label source-reference objective: its expected source-relative log-risk remainder at 32 labels is −log⁡(0.8)/33=0.006762-\log(0.8)/33=0.006762 nats. Theorem 1 states the individual-candidate oracle explicitly, and Appendix A.6 gives the counterexample distinguishing it from the convex-hull oracle. We also pursue the stronger objective with a separate continuous-weight estimator, TL-CBMA. On the exact three-candidate mixture target, its prefix improves NLL over the indexed endpoint by 0.111 nats; MAP4 and MLE are still lower by 0.00321 and 0.0333 nats. Its best-mixture remainder is 0.242 nats for three candidates and 4.14 nats for 774 at the same budget. These results explain how the choice of oracle affects the estimator and its finite-budget cost (Appendix E.1; Table 32).

W5. Does the construction have uses beyond the projected Burgers problem?

The new applications share the same purpose: retain a useful task forecast while learning the joint information it leaves open. Weather J121 adapts gusts while retaining temperature; hydrology adapts flow shape under a volume forecast; advection–diffusion J10 improves off-axis full-field CRPS by 1.08% while retaining total mass. In six-particle dynamics, J81 improves joint NLL over symmetric, random-direction and normal-score comparators under the original Holm comparison family, while the physics-informed baseline remains stronger. These application-specific dictionaries are identified separately from Burgers C774 (Section 5.4; Tables 23–31).

AuthorsOpenReview ↗

Response to Reviewer AGNZ (2/2): Answers to Your Questions

These answers continue the source-mechanism argument through physical prediction and advance selection. References are to the revised manuscript; improvements are reductions in the named loss.

Q1. How is a source-selected direction useful when the target shift is unknown?

The direction represents a specified forcing-response mechanism. We now test where that mechanism remains useful on new target fields. It captures 99.77% and 99.80% of the measured conditional-mean slope change under the two correlation controls, versus 15.56% under the preselected outside control. Under negative correlation, original C774 lowers region-2 CRPS by 8.71×10−58.71\times10^{-5}, with improvement interval [4.20,13.2]×10−5[4.20,13.2]\times10^{-5}. These target moments explain the observed outcome after evaluation; the deployed rule uses the source direction. Independent source checks reproduce its selected pair in 20/20 cases, with 16/20 agreement with the full nonlinear covariance-response criterion.

We do not yet have a reliable pre-deployment rule for shifts caused by previously unseen mechanisms. The evidence locates the useful mechanism, and the earlier unknown-pair challenge is retained with its interval crossing zero (Sections 3.1 and 5.3; Table 21).

Q2. How does a misspecified task marginal affect joint prediction?

It contributes an explicit additive task-error term. For the pooled conditional, the joint KL risk decomposes as

EXKL(PZU∣X∥mDHD)=EXKL(PZ∣X∥mD)+I(U;X∣Z)+A(HD).\mathbb E_X\mathrm{KL}(P_{ZU\mid X}\Vert m_DH_D)=\mathbb E_X\mathrm{KL}(P_{Z\mid X}\Vert m_D)+I(U;X\mid Z)+A(H_D).

With a common marginal, coordinates and completion, conditional and joint log-risk differences are equal, including when both fitted modules use the same target labels (Section 4.1; Corollary 1). Weather illustrates both interventions. With the temperature marginal fixed, J121-prefix adaptation lowers joint NLL from 4.525 to 4.253. With the pooled source conditional fixed, an independent 2017 temperature update lowers task and joint NLL by the same 0.0349 nats (Table 25). Thus an upstream update and conditional adaptation have separately measurable contributions.

Q3. Do coefficient-level gains reach physical outputs?

They do in identified settings, which we now evaluate directly. Original C774 lowers G5 tail OR-event Brier by 3.31×10−43.31\times10^{-4}, with improvement interval [1.31,5.30]×10−4[1.31,5.30]\times10^{-4}. The negative-correlation CRPS gain in Q1 provides a second physical outcome. The earlier F2 tail experiment also lowers region-2 CRPS by 2.63×10−42.63\times10^{-4}, with interval [0.320,4.95]×10−4[0.320,4.95]\times10^{-4}. Under G5 combined shift, however, NLL improves by 0.215 nats while region-2 CRPS rises by 1.18×10−41.18\times10^{-4}. This is why the revision scores the physical output alongside likelihood. In the separate J10 advection–diffusion experiment, off-axis full-field CRPS improves by 1.08%, while mass NLL and CRPS stay unchanged (Tables 1, 15 and 30).

Q4. How sensitive are gains to the prior and grid?

More reference mass strengthens source protection while leaving substantial adaptation gains. Moving the reference mass from 0.5 to 0.95 lowers null-shift cost from 0.00681 to 5.10×10−55.10\times10^{-5} nats; tail and combined gains remain 0.296 and 0.334 nats. The study covers four priors, two nested radial grids, six controls and three estimators. Reducing the radial grid from 17 to 9 nodes changes mean prefix NLL by at most 2.01×10−42.01\times10^{-4} nats (Tables 18–20). The expected bound −log⁡πref/(K+1)-\log\pi_{\rm ref}/(K+1) gives the corresponding source-protection interpretation.

Q5. Which events benefit, and can we predict the benefit in advance?

The tail OR-event gain above is one direct C774 example. For advance selection, we froze a policy that estimates event frequency from calibration features and predicts Brier gain as

g^e=(p1e−p0e)(2η^e−p0e−p1e).\widehat g_e=(p_{1e}-p_{0e})(2\widehat\eta_e-p_{0e}-p_{1e}).

It selects the adapted branch for both AND/OR events when their mean predicted gain is positive. Over the seven-control benchmark, whole-joint selection lowers mean Brier by 1.30×10−41.30\times10^{-4} versus always-source and 4.21×10−44.21\times10^{-4} versus always-adapt, improving all four source means. The respective descriptive intervals are [1.20,1.40]×10−4[1.20,1.40]\times10^{-4} and [2.34,6.08]×10−4[2.34,6.08]\times10^{-4}. This policy selects between source m/Hm/H and CV32 mm/C774 prefix, with common completion. H-only selection keeps CV32 mm fixed and has smaller differences whose intervals cross zero. The separate Clopper–Pearson rule accepts 1 of 448 whole-joint contexts, rejects 0 and leaves 447 uncertain. The point-policy improvement and confidence-classification coverage therefore answer different parts of the advance-selection question (Table 3; Appendix E; Table 33).

thank u but retain

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author response: Summary of revisions

The revised manuscript follows one argument: when an application relies on an existing task forecast, a small target sample can adapt the remaining joint structure while preserving that forecast. The original C774 estimator and its expected source-reference guarantee remain central. The revision connects this construction to quantitative representation design and measured predictive benefits.

We clarify how source data define the conditional family and how target labels update its weights. Figure 1 illustrates this interface. The symmetry analysis then identifies which changes the family can represent, while matched direction and covariance interventions show how representation and label budget interact.

The empirical comparisons follow the same logic. Original C774 improves conditional NLL by 0.4503 nats under the original Burgers combined shift. Same-family source-regularized fits and a matched conditional BNN clarify the estimator comparison; Figure 2 presents the main G5 contrasts. Against the BNN selected on separate development data, C774 has a mean NLL advantage of 0.0968 nats in the supplemental G5 comparison.

Weather, hydrology, advection–diffusion and particle studies connect the construction to physical predictions. In hydrology, retaining the supplied volume forecast reduces volume CRPS by 26.3% relative to an additional sparse update with conditional shape fixed. Absolute scores and calibration diagnostics also assess the supplied task forecasts directly.

The continuous-weight estimator TL-CBMA and frozen branch-selection policy address further questions raised in review as separate extensions. Our reviewer-specific responses explain the comparisons and point to the corresponding sections and tables in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.