acceptodds
Under review as a conference paper at ICLR 2027

From Fitting to Flowing: Reliability-Guided Geometric Correction Dynamics for Multi-View Partial Multi-Label Learning

Abstract

Multi-view partial multi-label learning (MVPML) learns from heterogeneous views and ambiguous candidate label sets containing both relevant and irrelevant labels. Existing methods typically refine candidate supervision or jointly optimize disambiguation and classification, but ultimately treat inferred supervision as a target to fit. Errors in this supervision can therefore propagate directly into prediction, while multi-view geometry is mainly used for refinement or regularization rather than for controlling such propagation. We propose GeoCURVE, a reliability-guided, geometry-embedded continuous-time dynamical framework that instead treats inferred supervision as a controlled correction to a structural reference. Specifically, prediction evolves according to a reference-restored ordinary differential equation (ODE). Candidate reliability, indicating the likelihood that a candidate label is truly relevant, determines the correction signal, shared multi-view geometry governs how it propagates across structurally related instances, and a restoration term controls how much of it is retained. The resulting dynamics admit a unique globally exponentially stable equilibrium with a closed-form solution, exposing how reliability, geometry, and restoration jointly determine the final prediction.We derive an oracle-based error bound quantifying controlled uncertainty propagation and an incremental generalization bound showing how spectral geometry and restoration control correction complexity and the correction-induced generalization gap. Experiments on six benchmark datasets under multiple ambiguity settings demonstrate consistent improvements over representative MVPML methods, achieving relative AP improvements of up to 13% over competitive baselines on several datasets, while ablations show that controlling correction extent, rather than merely refining supervision, drives the gain.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-View Partial Multi-Label Learning, Continuous-Time Dynamics, Candidate-Label Reliability, Multi-View Geometry

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_R5pEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the problem of multi-view partial multi-label learning, where each sample is described by multiple views, and the candidate label set comprises all true labels alongside some false ones. The authors propose GeoCURVE, which utilizes reliability-weighted candidate labels as correction signals for a reference prediction, while controlling the extent of correction through a shared kernel structure and a recovery term.

Specifically, the method first fuses Gaussian kernels from various views to construct a reference prediction via kernel ridge regression. It then estimates candidate label reliability using leave-one-out prediction scores and a density ratio model for trustworthy negative labels, thereby generating soft supervision and the corresponding residuals relative to the reference prediction. The prediction is formulated as the state of a continuous-time system driven by the shared kernel toward the soft supervision, while simultaneously being pulled back toward the reference prediction by the recovery term. Predictions are ultimately derived from the system's closed-form equilibrium solution and extended to new samples via cross-kernel mapping. Theoretical analysis covers dynamical stability, supervision error propagation, and correction complexity; experiments validate the method across six datasets and various label ambiguity configurations, with ablation studies and mechanism diagnostics analyzing the roles of individual components.

Critical strengths

1.The mathematical formulation is clear, and the theoretical analysis and closed-form prediction framework are comprehensive.

The paper explicitly delineates the relationships among reference prediction, reliability estimation, residual correction, and the final balanced solution. Theoretical analysis encompasses dynamic stability, supervised error propagation, and the complexity of the correction function; it also explains—from a spectral perspective—how geometry and recovery strength influence the correction process. The closed-form solution renders the roles of reference retention and supervised correction intuitive, avoids numerical integration, and provides a framework for inductive prediction on unseen samples. These analyses facilitate an understanding of the method's behavior and its conditions of applicability.

2.The experimental results and diagnostic analyses are extensive.

The experiments cover six datasets, six label ambiguity configurations, seven competing methods, and five evaluation metrics. The paper reports that the proposed method achieves the best performance in 33 out of 36 settings (based on average precision) and presents further ablation and parameter sensitivity analyses. Beyond final performance, the authors examine the correlation between reliability scores and ground-truth labels, the distribution of evidence regarding erroneous candidates versus trustworthy negative labels, and the alignment between the correction direction and oracle corrections. These diagnostic insights provide supporting evidence for the intended mechanisms of reliability estimation and controlled correction.

Critical weaknesses

1.The extra value of the dynamical formulation over static regularization models remains unclear

While the dynamics provide an intuitive explanation for the correction and recovery processes, the paper actually employs a closed-form equilibrium solution rather than utilizing finite-time evolution. Let f0f_0 be the reference function; consider the static objective:

min⁡h∈HkC∑i=1n∥f0(xi)+h(xi)−Y~i∥22+λc∥h∥HkC2,\min_{h\in\mathcal{H}_k^C} \sum_{i=1}^{n} \left\|f_0(x_i)+h(x_i)-\widetilde{Y}_i\right\|_2^2 +\lambda_c\|h\|_{\mathcal{H}_k^C}^2,

The solution is

h(x)=kx⊤(K+λcI)−1(Y~−P0),h(x)=k_x^\top(K+\lambda_c I)^{-1} (\widetilde{Y}-P_0),

which aligns with the paper's formulation for correction and inductive prediction. Consequently, the final estimator can also be interpreted as regularized kernel regression centered around the reference function.

This equivalence does not negate the value of the dynamical interpretation, but it necessitates further clarification regarding the contribution emphasized by the paper—specifically, the shift "from fitting to evolution." The current analysis has not sufficiently demonstrated what modeling capabilities, theoretical insights, or empirical benefits the dynamical approach offers that static models cannot easily provide. In particular, stable convergence and constrained correction do not directly imply superior predictive performance compared to other appropriately regularized estimators.

2.The handling of view discrepancies in multi-view fusion is limited, lacking verification of associated benefits and robustness.

Multi-view information is incorporated into the model primarily through the equal-weight summation of kernels from each view. While this approach aggregates sample relationships across views, it fails to explicitly distinguish view reliability or specifically model the divergent or complementary relationships between views. When certain views contain noise or provide conflicting similarity relationships, equal-weight fusion may dilute useful structural information, thereby compromising the reference prediction, candidate reliability estimation, and the final correction.

Since the fused prediction process relies solely on a shared kernel, the extent to which the method actually leverages information unique to the multi-view setting requires further differentiation. Existing experiments lack direct comparisons with approaches such as single-view methods, concatenated feature kernels, and adaptive weighted kernels; furthermore, scenarios involving uneven view quality or structural conflicts have not been adequately investigated. This limits the assessment of the necessity and applicability of the proposed multi-view mechanism.

Questions for authors

1.Contributions regarding dynamics

Do the authors acknowledge the equivalence between the aforementioned reference-centered regularized kernel regression and the proposed equilibrium predictor? Under this equivalence, what modeling capabilities or theoretical insights does the dynamical formulation offer that are difficult to obtain through static expressions? Given that the experiments directly utilize equilibrium solutions, which results serve to distinguish the contributions of the dynamical perspective from those of reference construction, reliability estimation, and standard regularization?

2.Regarding multi-view fusion Under comparable model selection conditions, what are the gains of equal-weight shared kernels compared to single-view kernels, concatenated feature kernels, and adaptive weighted kernels? When view quality is unbalanced or sample relationships conflict, how does the method prevent unreliable views from simultaneously affecting both the reference prediction and the reliability estimation? Could experiments involving controlled view perturbations be provided to verify the robustness of this fusion mechanism?

1 comment

Thank you for the detailed review.

1. Additional value of the dynamical formulation (CW1, Q1). Thank you for this insightful comment. We acknowledge that GeoCURVE's equilibrium predictor is equivalent to a reference-centered residual KRR estimator. The contribution of the ODE is therefore not to introduce a different equilibrium estimator, but to characterize the evolution of correction from the structural reference toward equilibrium.

For 0<τ=ξct<∞0<\tau=\xi_c t<\infty and μ>0\mu>0, the finite-time spectral response is gτ(μ)=μμ+λc[1−e−τ(1+μ/λc)]g_\tau(\mu)=\frac{\mu}{\mu+\lambda_c}[1-e^{-\tau(1+\mu/\lambda_c)}], corresponding to an eigenvalue-dependent effective ridge parameter λeff,τ(μ)=λc+μ+λceτ(1+μ/λc)−1\lambda_{\mathrm{eff},\tau}(\mu)=\lambda_c+\frac{\mu+\lambda_c}{e^{\tau(1+\mu/\lambda_c)}-1}. Thus, finite-time correction generally cannot be reproduced by a conventional residual KRR with a single global regularization parameter.

The continuous-time dynamics reveal a trade-off between reducing reference bias and controlling correction-induced uncertainty and complexity. Partial correction limits the influence of uncertain inferred supervision and correction complexity, whereas more extensive correction reduces reference bias at an increasing statistical cost. As a result, an intermediate correction level may achieve a more favorable balance than equilibrium correction. We have revised Theorems 1 and 2 to characterize this trade-off through finite-time uncertainty control, trajectory-wide generalization bounds, and a sufficient condition for finite-time risk improvement.

To isolate the dynamical contribution, our additional trajectory experiments (see Sec. 5.5) vary only the evolution time while keeping the shared kernel, structural reference, inferred supervision, and regularization parameters fixed. The resulting trajectories exhibit non-monotonic, dataset- and ambiguity-dependent performance variations, with finite-time optima observed in many configurations. These results demonstrate the predictive benefits of partial correction beyond the effects of reference construction and reliability estimation.

We retain the closed-form equilibrium in Algorithm 1 to avoid additional stopping-time selection and its associated model-selection cost, while maintaining competitive prediction performance.

We have clarified the equilibrium equivalence to residual KRR, revised Theorems 1 and 2, and added controlled finite-time trajectory experiments to characterize the theoretical and empirical contributions of dynamical correction in the revised manuscript.

2. Multi-view fusion and robustness (CW2, Q2). Thank you for this insightful comment. We have conducted additional comparisons with three alternative fusion strategies: single-view kernels, concatenated-feature kernels, and prediction-discrepancy-based adaptive-weight kernels. The results show that equal-weight fusion achieves an average AP of 0.6642, substantially outperforming the single-view baseline (0.6317) while remaining comparable to feature concatenation (0.6648) and adaptive fusion (0.6640). It also outperforms the single-view baseline on five of the six datasets, demonstrating the benefit of aggregating complementary information across views. Notably, equal-weight fusion achieves performance comparable to adaptive fusion without the additional computation required for view-specific weight estimation, making it a simple and efficient fusion strategy.

We further conduct controlled perturbation experiments by injecting Gaussian noise into individual views and randomly permuting sample correspondences. The results show that multi-view fusion consistently exhibits smaller performance degradation than the single-view baseline under both perturbations. Adaptive weighting achieves the smallest degradation by reducing the contributions of views with larger prediction discrepancies. However, the improvements over equal-weight fusion are modest, and the two strategies maintain comparable predictive performance. This indicates that equal-weight fusion provides a favorable balance between robustness and computational simplicity, without requiring additional view-specific weight estimation. We also examine the effects of view perturbations on reference prediction and candidate-reliability estimation.

Overall, these results support the effectiveness and empirical robustness of shared multi-view kernel fusion, while demonstrating the competitiveness of the simple equal-weight strategy without additional view-weight estimation.

We will include these additional experimental results and the corresponding discussions in the revised manuscript.

Reviewer_5eKvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoCURVE, a reliability-guided geometric correction framework. The method constructs a kernel ridge regression reference, estimates candidate-label reliability using leave-one-out predictions and a trusted-negative density-ratio model, and then, introduces reference-restored continuous-time dynamics whose equilibrium provides the final prediction. The main contribution can be described as a controlled-correction formulation with explicit spectral regulation and accompanying uncertainty-propagation and incremental generalization bounds. Experiments on six benchmarks show the best average precision in 33 of 36 dataset–ambiguity configurations, with ablations demonstrating consistent benefits from reference restoration.

Critical strengths
  1. The paper addresses a relevant issue: refined candidate supervision can remain erroneous, so its influence on prediction should be regulated. The formulation separates reference construction, reliability estimation, geometric propagation, and restoration, making their respective roles explicit.

  2. The dynamics admit a closed-form trajectory and a unique globally exponentially stable equilibrium. The spectral characterization explains how kernel eigenvalues and restoration strength determine correction retention, while Theorems 1 and 2 quantify supervision-error propagation and correction complexity.

  3. GeoCURVE achieves the highest AP in 33 of 36 configurations and remains among the top two in every configuration.

  4. The reliability analysis additionally reports improved oracle correction alignment in 223 of 240 outer folds. the Appendix also specifies that preprocessing and hyperparameter selection are confined to training splits and use weak supervision.

Critical weaknesses
  1. From Eqs. (12)–(13), the implemented predictor is a KRR reference plus a second KRR predictor fitted to the residual Rcs=Y~−P0R_{\mathrm{cs}}=\widetilde{Y}-P_0. In particular, the correction is also the solution of the static objective $ \min_{h\in\mathcal{H}k^C} |h(X)-R{\mathrm{cs}}|F^2+\lambda_c|h|{\mathcal{H}_k^C}^2. $ The Appendix confirms that prediction uses the equilibrium without numerical integration. Thus, although the dynamical interpretation is informative, the paper lacks discussion of this equivalence and should clarify what the ODE perspective contributes beyond the resulting static estimator. Comparisons with direct KRR on reliability-informed supervision and simpler regularized correction schemes would help establish the distinctive contribution.

  2. Section 5 and Appendix describe nested cross-validation over 105 capacity pairs for GeoCURVE. In contrast, Appendix D.2 lists fixed recommended settings for most competitors, including GRADIS, FIMAN, GLADE, MSPML-TLR, and VADIS. Although a search grid is listed for TMvML, its selection procedure is not explained there. Im thinking that the reported gains may partly reflect differences in tuning effort.

  3. Table 2 shows that the geometry-free variant exceeds the full model on EspGame5k at ((p,r)=(0.3,1)(p,r)=(0.3,1)), ((0.3,3)(0.3,3)), and ((0.7,1)(0.7,1)), with AP values of 0.456 versus 0.453, 0.450 versus 0.448, and 0.453 versus 0.450. The reliability-free variant also matches or exceeds the full model in several configurations. Section 5.4 appropriately acknowledges these configuration-dependent effects, but all variants reuse capacities selected for the full model.

  4. Eq. (4) approximates false-candidate evidence using the trusted-negative distribution, while Appendix B4 additionally requires the true-candidate contribution to become negligible in the low-evidence region. Figures 2 & 5 show that false candidates are closer to trusted negatives than true candidates in JS divergence, yet, this relative similarity does not establish either distributional equivalence or the required tail condition.

  5. The paper should clarify the gap between the theoretical assumptions and implementation. Theorem 2 explicitly assumes that the kernel, including preprocessing and bandwidths, is fixed independently of the sample. However, Appendix D3 estimates feature means, standard deviations, and RBF bandwidths from the training sample. Appendix C2 handles data-dependent reference and correction functions through a uniform concentration event, but its argument still relies on a fixed feature map.

Questions for authors
  1. Eqs. (12)–(13) appear to make the final correction equivalent to kernel ridge regression on the residual (Rcs=Y~−P0R_{\mathrm{cs}}=\widetilde{Y}-P_0), and Appendix D3 states that the equilibrium is computed directly without numerical integration. Do you agree with this static interpretation?

  2. GeoCURVE selects ((γr,γc)(\gamma_r,\gamma_c)) through nested cross-validation over 105 parameter pairs, whereas Appendix D2 appears to use fixed recommended settings for most competing methods. Were the baseline hyperparameters also selected separately within each outer-training fold using weak supervision?

  3. Table 2 reuses the capacity parameters selected for the full model for all ablation variants. Since removing a component changes the estimator and may change its preferred regularization, could the authors report whether the conclusions remain the same when each ablation is independently tuned?

  4. Theorem 2 assumes that the kernel, including its preprocessing and bandwidths, is fixed independently of the sample. However, Appendix D3 estimates feature-standardization statistics and RBF bandwidths from each training split. Does Theorem 2 apply to this sample-dependent kernel construction, for example through an appropriate conditional argument, or should the theorem be understood as covering a different setting from the reported implementation?

1 comment

Thanks for the constructive comments.

1. Equilibrium residual KRR and the contribution of dynamics (CW1, Q1). We acknowledge that the equilibrium correction is exactly the solution of residual KRR centered at the structural reference P0\mathbf P_0. Direct KRR on Y~\widetilde{\mathbf Y} is different, as it does not preserve P0\mathbf P_0 as the correction reference.

The ODE additionally characterizes the correction trajectory. At finite time, its spectral response is gτ(μ)=μμ+λc[1−e−τ(1+μ/λc)](τ=ξct)g_\tau(\mu)=\frac{\mu}{\mu+\lambda_c}[1-e^{-\tau(1+\mu/\lambda_c)}] (\tau=\xi_c t), corresponding to an eigenvalue-dependent effective ridge parameter λeff,τ(μ)=λc+μ+λceτ(1+μ/λc)−1\lambda_{\mathrm{eff},\tau}(\mu)=\lambda_c+\frac{\mu+\lambda_c}{e^{\tau(1+\mu/\lambda_c)}-1}. Thus, finite-time correction generally cannot be reproduced by a single global ridge parameter.

The continuous-time dynamics reveal a trade-off whereby partial correction limits the influence of uncertain inferred supervision and correction complexity, whereas more extensive correction reduces reference bias at an increasing statistical cost. As time increases, reference bias decreases, while the influence of inferred-supervision error and correction complexity grows, suggesting that partial correction may achieve a better balance than equilibrium correction. We have revised Theorems 1 and 2 to characterize this trade-off and establish conditions for finite-time risk improvement, and added trajectory experiments (see Sec. 5.5) showing non-monotonic performance with finite-time optima in many configurations. We retain the closed-form equilibrium to avoid additional stopping-time selection while maintaining competitive prediction performance.

2. Hyperparameter selection and tuning fairness (CW2, Q2). Most baselines use recommended hyperparameters or official implementations, which provide established optimal and robust configurations. For TMvML, we use the official implementation, which selects hyperparameters using ground-truth labels. In the revised version, we further evaluate representative parameter settings under representative ambiguity configurations across datasets. The additional results do not change the overall comparative conclusions.

3. Ablation study and independent hyperparameter tuning (CW3, Q3). We reran the ablation experiments, independently tuning each variant using the same nested cross-validation protocol, parameter grids, and weak-supervision-based SoftAP criterion as the full model. We corrected the geometry-free ablation setting by using a normalized constant kernel for both the training kernel and the test–training cross-kernel for a fair comparison. The updated results remain consistent with the original conclusions: removing restoration reduces AP in all 24 configurations, while removing correction generally degrades performance, particularly on Yeast, OBJECT, and Human. The configuration-dependent effects of LOO, reliability, and geometry further highlight the complementary roles of these components. Overall, GeoCURVE's coordinated design performs favorably among the independently tuned variants. Full results are reported in the revised manuscript.

4. Negative-like mixture approximation and low-evidence tail condition (CW4). We supplemented JS divergence analysis with cumulative tail diagnostics. Across 240 cases (six datasets, four ambiguity settings, ten folds), the average divergences are DJS(F,N)=0.0011D_{\mathrm{JS}}(F,N)=0.0011, DJS(T,N)=0.3260D_{\mathrm{JS}}(T,N)=0.3260, and DJS(F,T)=0.3232D_{\mathrm{JS}}(F,T)=0.3232, with both inequalities holding in all cases. These results show that false-candidate evidence closely resembles trusted negatives, unlike true-candidate evidence. As the threshold decreases from 0.500.50 to 0.050.05, GT(ζ)/GN(ζ)G_T(\zeta)/G_N(\zeta) decreases from 0.1800.180 to 0.0910.091, while GF(ζ)/GN(ζ)G_F(\zeta)/G_N(\zeta) remains near one. These ratios compare the low-evidence mass of true and false candidates with trusted negatives, indicating diminishing true-candidate contribution and similar tail behavior between false candidates and trusted negatives. Together, these findings empirical support negative-like approximation and the assumed low-evidence tail behavior.

5. Sample-dependent kernel construction and theoretical consistency (CW5, Q4). We have extended Theorem 2 to accommodate the training-estimated preprocessing and RBF bandwidths used in our implementation. Specifically, we introduce a fixed reference feature map and establish uniform concentration over the bandwidth family, with a feature-displacement bound accounting for data-dependent preprocessing. This aligns the theoretical analysis with the actual kernel construction. Importantly, under sample-independent preprocessing and bandwidths, the additional covering and displacement terms are unnecessary, and the equilibrium bound recovers the original incremental generalization bound. Thus, the revised theory extends the original result while preserving its fixed-kernel guarantee.

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

This paper studies multi-view partial multi-label learning and proposes GeoCURVE. The method fuses view-specific RBF kernels, constructs a reference prediction through kernel ridge regression, and estimates candidate-label reliability using leave-one-out predictions and a density ratio based on trusted negatives. The difference between reliability-weighted supervision and the reference prediction forms a correction signal, regulated by a linear ODE with kernel geometry and a reference-restoration term. Prediction uses the closed-form equilibrium and a cross-kernel extension. The authors provide analyses of stability, supervision-error propagation, and incremental generalization, together with comparisons, ablations, and reliability diagnostics on six datasets under six label-ambiguity settings.

Critical strengths
  1. The reference prediction, reliability estimation, and correction control fit together clearly. Leave-one-out prediction reduces labels' direct influence on their own reliability scores, and the closed-form solution avoids numerical integration. Theorem 1 separates reference contamination from inferred-supervision error, helping explain how regularization affects uncertainty propagation.

  2. Table 1 reports the best AP in 33 of the 36 dataset–setting combinations, with inner-loop tuning using only weak supervision. Diagnostics relating reliability to true relevance and examining correction direction provide further support for the method's practical value.

Critical weaknesses
  1. Eqs. (12)–(13) are equivalent to residual KRR with target Y~−P0\widetilde{Y}-P_0, and Algorithm 1 uses only the equilibrium. The added contribution of the dynamical formulation therefore remains unclear. Removing restoration in Table 2 removes regularization of the correction, so the resulting performance drop does not by itself distinguish a dynamical contribution from ordinary shrinkage.

  2. Appendix D.5 treats six ambiguity settings on each of six datasets as N=36N=36 testing units, although settings on the same dataset share samples. Ignoring this dependence may overstate significance. The existing results should be analyzed with grouping by dataset, or the claims narrowed; this does not invalidate the descriptive advantage.

Questions for authors
  1. Compared with the equivalent residual KRR formulation with a reference prior, what additional algorithmic design or theoretical insight does the continuous-time formulation provide? Please explain using the existing analysis.

  2. How do the existing results in Table 2 distinguish the contribution of the correction mechanism from the benefits of the reference prediction itself and ordinary regularization shrinkage?

  3. Which significance conclusions remain if ambiguity settings are aggregated with datasets as the independent units, or analyzed using a method that accounts for within-dataset dependence? Please explain using the existing results.

2 comments

Thank you for the constructive review. We have revised the theoretical analysis, added trajectory experiments, and re-examined the statistical results.

1. Dynamical formulation beyond equilibrium residual KRR (CW1, Q1). We acknowledge that GeoCURVE's equilibrium predictor is equivalent to residual kernel ridge regression (KRR) centered at the structural reference. The ODE contribution is therefore not a different equilibrium estimator, but a characterization of the correction trajectory and the benefits of partial correction.

For 0<τ=ξct<∞0<\tau=\xi_c t<\infty and μ>0\mu>0, the finite-time spectral response is gτ(μ)=μμ+λc[1−e−τ(1+μ/λc)]g_\tau(\mu)=\frac{\mu}{\mu+\lambda_c} [1-e^{-\tau(1+\mu/\lambda_c)}], with the eigenvalue-dependent effective ridge parameter λeff,τ(μ)=λc+μ+λceτ(1+μ/λc)−1\lambda_{\mathrm{eff},\tau}(\mu) =\lambda_c+\frac{\mu+\lambda_c} {e^{\tau(1+\mu/\lambda_c)}-1}. Hence, when the residual has components in multiple distinct positive eigendirections, finite-time correction cannot generally be reproduced by a single global ridge parameter.

The continuous-time formulation reveals a trade-off between reducing reference bias and controlling correction-induced uncertainty and complexity. At earlier times, partial correction retains more reference bias but limits the contribution of inferred-supervision error and correction complexity. As time increases, reference bias and contamination are progressively reduced, at the cost of greater reliance on inferred supervision and increased correction complexity. Thus, an intermediate correction level may achieve a more favorable balance than equilibrium correction. To characterize this trade-off, we have revised the theoretical analysis: Theorem 1 establishes the time-dependent control of supervision uncertainty, while Theorem 2 characterizes the evolution of correction complexity and generalization bounds and provides a sufficient condition for finite-time risk improvement. We further added trajectory experiments, which reveal non-monotonic, dataset- and ambiguity-dependent performance variations, with finite-time optima observed in many configurations. We have also clarified this equilibrium equivalence and discussed the distinction between finite-time dynamics and static residual KRR in the revised manuscript.

We nevertheless retain the closed-form equilibrium in Algorithm 1 to avoid an additional stopping-time hyperparameter and model-selection cost, while maintaining competitive prediction performance.

2. Contributions of reference prediction and correction mechanism (CW1, Q2). The w/o Correction variant retains only the structural reference P0\mathbf P_0, its comparison with the full model therefore measures the additional benefit of candidate-specific correction. The w/o Restoration variant removes the reference-restoring constraint, and its consistently lower AP across the reported configurations highlights the importance of controlling correction extent. Since the ablations reuse the capacities selected for the full model, they assess component contributions under a common setting rather than independently optimized variants. These ablations do not, by themselves, isolate the dynamical contribution from single-parameter KRR shrinkage. The finite-time spectral analysis further shows that its response cannot generally be reproduced by a single global ridge parameter, while trajectory experiments illustrate its performance implications.

We have clarified the interpretation of the ablation results in the revised manuscript.

3. Statistical significance and within-dataset dependence (CW2, Q3). We averaged the six ambiguity configurations within each dataset and treated the six datasets as independent units (N=6N=6, K=8K=8). We then applied the Friedman test with the Iman--Davenport correction and the Bonferroni--Dunn post-hoc test, using GeoCURVE as the control.

The null hypothesis of equal performance is rejected for all five metrics at α=0.05\alpha=0.05. GeoCURVE obtains the best average rank on AP, OE, RL, and COV, and the second-best on HL. The post-hoc comparisons are:

MetricFFF_FCritical valueGeoCURVE rankNot significantly different
AP11.29312.28521.167GRADIS, FIMAN, TMvML, FastGRAIL
HL37.00002.28522.000GRADIS, FIMAN, VADIS, TMvML, FastGRAIL
OE12.26032.28521.000GRADIS, FIMAN
RL10.79942.28521.167GRADIS, TMvML, FastGRAIL
COV10.61982.28521.167GRADIS, TMvML, FastGRAIL

We have replaced the previous statistical analysis with the dataset-level results to account for within-dataset dependence in the revised manuscript.

I thank the authors for the thorough rebuttal. The dataset-level significance analysis fully addresses CW2. However, CW1 is explained rather than resolved: Algorithm 1 still uses only the equilibrium, so the dynamical formulation contributes insight without changing the estimator. While avoiding an extra stopping-time hyperparameter is a practical choice, the dynamical framework's benefits remain analytical rather than algorithmic. My score remains unchanged.

Comments on the paper

1 comment

Dear Reviewers and Authors,

As a quick reminder, the author-reviewer discussion period will close on November 18.

@Reviewers: Please make sure to read the authors' rebuttals, acknowledge their responses, and engage in further discussion if any points remain unclear. While final decisions are not made at this stage, we ask that you carefully re-evaluate your ratings in light of the new clarifications and experimental results:

Increase your score if the authors have satisfactorily resolved your concerns and strengthened the paper.

Decrease your score if new, unaddressed issues have emerged during the discussion.

Maintain your score if critical concerns remain unresolved.

Please note that you are under no obligation to change your rating, provided your final score accurately reflects the current state of the manuscript. To ensure a fair evaluation, please ground your assessment in concrete evidence rather than personal preference, and consider whether the work offers meaningful scientific value to the ICLR community.

Thank you very much for your time, thoughtful engagement, and contributions to the review process.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.