acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Reliability Curriculum Guided Reinforcement Learning for Partial Label Disambiguation

Abstract

\em Partial-label learning (PLL) addresses the problem where each training sample is associated with a set of candidate labels, among which only one is the ground-truth label. The ambiguity of candidate labels brings great uncertainty to model optimization. Traditional disambiguation methods usually rely on classifier confidence or predefined label refinement rules. Such strategies may be affected by confirmation bias, leading to error accumulation and label noise propagation. To address these issues, this paper proposes a \em dynamic reliability curriculum guided reinforcement learning framework for PLL, called DC-RPLL. This method models label disambiguation as a Markov decision process, independently deciding for each sample whether to \em update, \em retain, or \em abstain. Specifically, DR-CPLD first represents each candidate label set as a soft label distribution, jointly uses classifier predictions and local feature space structure to construct disambiguation labels. Meanwhile, the reliability of each sample is estimated from neighborhood consistency, feature affinity, and soft label consistency to quantify the credibility of local structures. Based on this reliability, this paper designs dynamic curriculum thresholds so that highly reliable samples enter policy updates early, while low-reliability samples are gradually added later, thereby alleviating early error propagation. The Actor-Critic policy then guides policy learning according to the current confidence gain, neighborhood reliability, and label drift penalty, i.e., \em update, \em retain, and \em abstain. In addition, we develop a reliability-aware curriculum learning mechanism that gradually accepts samples from high to low reliability during training, effectively suppressing early error propagation. Experiments on benchmark and real datasets verify the effectiveness of the proposed algorithm.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Partial Label Disambiguation, Reinforcement Learning

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

This paper targets the confirmation bias and error accumulation issues caused by candidate label ambiguity in partial label learning, and proposes the DC-RPLL framework. The method estimates sample reliability using neighborhood information, filters samples from easy to hard via dynamic curriculum learning, then leverages Actor-Critic reinforcement learning to independently perform three label disambiguation actions for each sample, and designs a self-supervised reward function to realize policy optimization without ground-truth labels. The proposed method achieves better performance than multiple SOTA algorithms in experiments on benchmark datasets and real partial label datasets, and ablation experiments confirm that each module brings performance gains. However, this work lacks complete theoretical analysis, the model has non-negligible computational complexity, and further verification on larger-scale datasets is not conducted.

Critical strengths
  • Combining RL with sample reliability-aware curriculum learning to solve error propagation in partial label learning‌. Most existing partial label disambiguation methods update labels following static fixed rules, which easily amplifies errors generated in early training stages. This paper makes adaptive decisions of update, retain and abstain at the individual sample granularity, so as to dynamically avoid noisy label contamination introduced by low-reliability samples.

  • Designing a ground-truth-free self-supervised reward function for reinforcement learning policy training‌. Since no valid ground-truth labels are available in partial label scenarios, this reward function constructs supervision signals from four dimensions: confidence gain, entropy variation, neighborhood reliability and label drift, enabling the reinforcement learning paradigm to adapt to the weakly supervised setting.

Critical weaknesses
  • The theoretical analysis in this work is relatively weak. It only briefly mentions the error contraction property, and lacks rigorous derivations and proofs for the convergence of the reinforcement learning policy and the effectiveness of the self-supervised reward function. The core innovation of this paper lies in reinforcement learning-driven selective label disambiguation, but it does not fully demonstrate whether the policy can converge stably, or why the reward function can guide the model to obtain correct disambiguation results. The conclusions are only supported by experimental phenomena, which undermines the credibility of the method and stands as a critical shortcoming that affects the overall evaluation.

  • This framework needs to maintain neighborhood relationships and execute AC decisions for each individual sample, which leads to high time complexity. The paper only provides the complexity formula, without offering efficiency comparison results or feasible schemes to alleviate the computational burden.

Questions for authors
  • The paper mentions that the proposed reliability-aware correction has the error contraction property, but the appendix only provides the objective function analysis, without corresponding complete derivations. Could you supplement and specify under what exact assumptions this error contraction can be guaranteed to hold? If these assumptions cannot be satisfied, will DC-RPLL suffer from label disambiguation failure?

  • The self-supervised reward is the core of the entire reinforcement learning policy, and different hyperparameters will alter the weight of each reward term. Beyond the parameter sensitivity experiments presented in the paper, have you verified that when the reward weights are drastically perturbed, whether the policy will degenerate to always choosing abstention or performing blind updates, which further leads to severe performance degradation?

4 comments
AuthorsOpenReview ↗

We thank the reviewer for the careful review and valuable feedback. We address the concerns below.

We thank the reviewer for the careful review and valuable feedback. We address the concerns below.

Q1:

Thank you for identifying this gap. We have revised the appendix to provide complete derivations and explicit sufficient assumptions. The algorithm, reward function, correction target, and normalization order remain unchanged.

For the original affine correction, let Ui(e)U_i^{(e)} indicate an update action and define the current and proposal errors as di(e)=∥qi(e)−eyi∥1d_i^{(e)}=\|\boldsymbol q_i^{(e)}-\boldsymbol e_{y_i}\|_1 and d~i(e)=∥q~i(e)−eyi∥1\tilde d_i^{(e)}=\|\tilde{\boldsymbol q}_i^{(e)}-\boldsymbol e_{y_i}\|_1. The identity ∥q−ey∥1=2(1−q(y))\|\boldsymbol q-\boldsymbol e_y\|_1=2(1-q(y)) gives

=d_i^{(e)}+ U_i^{(e)}\beta_i^{(e)} (\tilde d_i^{(e)}-d_i^{(e)}).$$ Thus, a positive-step update reduces error exactly when the proposal assigns greater probability to the true class than the current label. Expected contraction requires: (i) feasible probability distributions with the stated candidate-support conditions; (ii) selected-proposal quality, $$\mathbb E[\tilde d_i^{(e)} \mid\mathcal F_e,U_i^{(e)}=1,\beta_i^{(e)}] \le\kappa d_i^{(e)}+\epsilon,\quad \kappa<1;$$ and (iii) effective update opportunities bounded by $$0<\omega_{\min}\le \mathbb E[U_i^{(e)}\beta_i^{(e)}\mid\mathcal F_e] \le\omega_{\max}.$$ These yield $$\mathbb E[d_i^{(e)}] \le\rho^e d_i^{(0)} +\frac{\omega_{\max}\epsilon}{1-\rho}(1-\rho^e), \qquad \rho=1-(1-\kappa)\omega_{\min}<1.$$ For the original operator that normalizes **after interpolation**, we preserve this order and derive its exact effective step: $$\lambda_i=\frac{\beta_i m_i}{1-\beta_i+\beta_i m_i}, \qquad m_i=\sum_{c\in S_i}\tilde q_i(c).$$ The corresponding proof uses $\lambda_i$ and the analytically conditioned proposal. We additionally relate proposal error to cross-class attention mass and same-class neighbors’ soft-label errors. High reliability alone does not establish these conditions. Zero residual error permits geometric contraction; persistent proposal error generally permits only a residual-error bound. If the assumptions fail, disambiguation does not necessarily fail, but contraction is no longer guaranteed. Incorrect consensus may reinforce errors, and persistent abstention may prevent recovery. Adaptive steps limit correction magnitude without certifying direction. Therefore, we have revised our claims regarding unconditional calibration and joint convergence, and we explicitly distinguish between the contraction of conditional label errors and the convergence of the overall training algorithm.

Q2:

Thank you for asking whether large perturbations to the self-supervised reward can cause persistent abstention or indiscriminate updates. To extend the sensitivity analysis in Fig. 4, we conducted a one-factor-at-a-time stress test on CIFAR-10 with q=0.3q=0.3. We varied the entropy-reduction weight λH\lambda_H, the reliability reward weight λR\lambda_R, and the label-drift penalty weight λD\lambda_D over 0,0.5,1,2,0, 0.5, 1, 2, and 1010 times their default values of 0.700.70, 0.250.25, and 0.100.10, respectively. The three sweeps share the same default configuration, giving 13 configurations in total. Each perturbation preserves the reward formulation and label-correction operator.

ConfigurationBest test accuracyUpdate coverage
DC-RPLL (default)89.15%95.62%
λH=0×\lambda_H=0\times87.91%92.50%
λH=0.5×\lambda_H=0.5\times88.87%92.76%
λH=2×\lambda_H=2\times88.41%92.27%
λH=10×\lambda_H=10\times87.47%92.80%
λR=0×\lambda_R=0\times86.91%92.41%
λR=0.5×\lambda_R=0.5\times88.40%92.89%
λR=2×\lambda_R=2\times87.10%92.76%
λR=10×\lambda_R=10\times85.99%92.54%
λD=0×\lambda_D=0\times87.54%92.75%
λD=0.5×\lambda_D=0.5\times88.21%93.10%
λD=2×\lambda_D=2\times87.44%92.60%
λD=10×\lambda_D=10\times87.32%92.30%

Based on these results, we can explicitly address the reviewer's two concerns:

  1. No Degeneration into "Always Abstaining": If the policy degenerated into always choosing abstention, the update coverage would plummet to near zero. However, as shown in the table, across all extreme perturbation settings,including completely zeroing out the entropy or reliability rewards,the update coverage remains highly stable, consistently staying above 92% (compared to 95.62% for the default). This demonstrates that the policy does not over-correct into a state of extreme conservatism (abstention).
  2. No Degeneration into "Blind Updates": If the policy engaged in blind updates, the introduction of excessive incorrect pseudo-labels would cause the accuracy to catastrophically collapse (e.g., dropping to near random guess or baseline levels without RL). While the accuracy does experience a controlled degradation under severe weight mismatch (e.g., dropping to 85.99% when λR=10×\lambda_R=10\times), it remains highly competitive and stable. The worst-case accuracy in this extreme perturbation matrix is 85.99%, which is still far above the random baseline and comparable to the non-learning contextual bandit baseline (86.11%) reported earlier. This indicates that the policy maintains its discriminative capability and does not blindly destroy the label space.

​ Over-emphasizing reliability (λR=10×\lambda_R=10\times) causes the most noticeable performance drop (85.99%), as it likely forces corrections on samples that are locally consistent but globally misclassified. Completely removing the entropy reward (λH=0×\lambda_H=0\times) degrades accuracy to 87.91%, proving its necessity in preventing premature convergence, yet the system remains stable. The distribution-drift penalty (λD\lambda_D) shows the highest tolerance to perturbation, with accuracy remaining robust across all scales. The extreme perturbation study confirms that while optimal reward balancing yields the highest accuracy (89.15%), the policy is highly robust against severe reward weight misspecification. It does not suffer from catastrophic collapse into permanent abstention or blind updating.

Reviewer_ZPydOpenReview ↗

Discussion

Thanks for your response.

Thank you for taking the time to review our response. We are happy to clarify any remaining questions. Best regards, The authors

Reviewer_tov4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DC-RPLL, which formulates label disambiguation as a Markov decision process and uses an Actor-Critic policy to select update, retain, or abstain actions. A reliability-aware curriculum prioritizes reliable samples to reduce early errors. Experiments on benchmark and real-world datasets demonstrate its effectiveness.

Critical strengths
  1. The research questions have practical value.
  2. The framework design has a certain degree of completeness.
  3. The report compares the proposed method with recent PLL methods, highlighting significant performance improvements in some areas, and provides component ablation and training dynamic visualization.
Critical weaknesses

The basis for reinforcement learning modeling needs to be supplemented.

Questions for authors
  1. Why is Actor-Critic necessary?
  2. Can the authors provide a controlled comparison of the individual mechanisms? In particular, can you isolate the effect of RL from reward design, curriculum masking, KNN propagation, and reliability-adaptive correction, and report incorrect-label-update rates alongside test accuracy?
2 comments
AuthorsOpenReview ↗

We thank the reviewer for the careful review and valuable feedback. We address the concerns below.

We thank the reviewer for the careful review and valuable feedback. We address the concerns below.

Q1:

The original w/o-AC ablation reported reduced performance when the learned policy was replaced by a fixed reliability rule. Meanwhile, we therefore conducted additional controlled experiments on CIFAR-10 with q=0.3q=0.3, keeping the backbone, candidate sets, reliability estimation, curriculum, correction operator, and execution-aligned reward identical. q=0.3$. We compared five decision strategies:

  • Reliability-based heuristic: Update when ri≥max⁡(ηe,0.8)r_i\ge\max(\eta_e,0.8); retain admitted instances below the additional update threshold; and abstain on non-admitted instances.
  • Greedy reward selection: Select the legal action with the highest observable immediate reward, without learning a policy.
  • Contextual bandit: A learned stochastic policy optimized using immediate action rewards without temporal-difference bootstrapping.
  • Actor–Critic with γ=0\gamma=0: Retain the sampled-action Actor–Critic optimization while removing future-value bootstrapping.
  • Full Actor–Critic with γ=0.9\gamma=0.9: Include the value of the next observation, collected at the subsequent refinement round after classifier training, in the TD target.
Reliability-based heuristic85.90%
Greedy reward selection85.76%
Contextual bandit86.33%
Actor–Critic with γ=0\gamma=086.01%
Full Actor–Critic with γ=0.9\gamma=0.989.15%

The key difference between γ=0\gamma=0 and γ=0.9\gamma=0.9 is whether the policy foresees the delayed penalties of a correction. Experimental results show that the complete Actor–Critic strategy achieves the best disambiguation performance, we take this as direct evidence of the gains from optimizing long-term rewards. Taken together, the original ablation and these matched comparisons support Actor–Critic as a consequential component of our framework, with future-value optimization providing an additional benefit in the evaluated setting. We will clarify that this is empirical support rather than a claim that Actor–Critic is universally necessary.

Q2:

On CIFAR-10 with q=0.3q=0.3, we compared Actor–Critic (AC) and a full-information contextual bandit under six configurations. Within each configuration, both strategies used the same backbone, candidate-label sets, state representation, actor architecture, reward definition, and correction operator. The bandit optimized the expected immediate reward over all legal actions. AC additionally used a learned estimate of future return.

ConfigurationAC best test accuracyAC IURBandit best test accuracyBandit IUR
Full configuration87.15%5.43%86.11%5.86%
Confidence-gain-only reward86.15%6.04%85.51%6.09%
Remove curriculum action mask only86.11%5.83%85.96%5.86%
Remove curriculum state feature only86.20%5.95%85.78%5.96%
Replace neighborhood correction proposal with classifier prediction85.99%7.53%85.23%7.25%
Use a common mean step size per refinement round86.09%6.00%85.98%6.05%

Best test accuracy is the highest value observed during training. The incorrect-label-update rate (IUR) is the proportion of actual updates whose post-update argmax differs from the true label. Each accuracy–IUR pair comes from the same run, with IUR accumulated over the entire training run.

Decision strategy and reward design. Under the full configuration, AC achieved 1.04 percentage points higher peak accuracy and 0.43 points lower IUR than the bandit. With the confidence-gain-only reward, the accuracy gap narrowed to 0.64 points. Relative to that simplified reward, the full reward increased peak accuracy by 1.00 point for AC and 0.60 points for the bandit, while reducing IUR by 0.61 and 0.23 points, respectively. The observed AC advantage therefore varied with the reward design. The AC–bandit comparison measures the difference between the two implemented decision strategies. A matched comparison between AC with γ=0\gamma=0 and γ=0.9\gamma=0.9 isolates future-value bootstrapping more directly.

Curriculum masking and state information. We removed the action mask while retaining the curriculum state feature, and separately zeroed that feature while retaining the mask. Both interventions lowered peak accuracy relative to the corresponding full configurations. Removing the mask increased IUR by 0.40 points for AC and left the bandit’s IUR unchanged at the reported precision. Removing the state feature increased IUR by 0.52 and 0.10 points, respectively. The admission logs are needed to determine how often the mask actually restricted updates and to interpret the differences in that ablation.

Neighborhood proposals and reliability-adaptive correction. Replacing the neighborhood correction proposal with classifier predictions produced the largest IUR increase for both strategies, 2.10 points for AC and 1.39 points for the bandit. This intervention retained neighborhood information in reliability estimation and the policy state, isolating its use in the correction proposal. The common-step-size control assigned every instance the mean of the proposed adaptive step sizes at each refinement round. It lowered peak accuracy and increased IUR for both strategies, with a larger accuracy decrease for AC (0.95 points) than for the bandit (0.13 points). These comparisons identify the neighborhood correction proposal as the component with the largest observed effect on incorrect updates in this setting.

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

This paper proposes DC-RPLL, a dynamic reliability curriculum guided reinforcement learning framework for partial-label learning. The method integrates KNN-based label propagation, instance-level reliability estimation, and curriculum learning to progressively refine ambiguous candidate-label distributions. An Actor-Critic policy selects among update, retain, and abstain actions, guided by a self-supervised reward incorporating confidence improvement, entropy reduction, reliability, and label drift. Experiments on four benchmark datasets and five real-world partial-label datasets demonstrate competitive performance against existing methods, while ablation studies and visualizations provide additional evidence of the framework's effectiveness.

Critical strengths
  1. Well-motivated selective disambiguation framework. The paper addresses confirmation bias and error propagation in partial-label learning through an interesting combination of reliability estimation, curriculum learning, and selective label correction. The ability to defer uncertain corrections is particularly appealing.

  2. The multi-faceted reward function (Equations 10-12) is well-designed for the PLL setting where true labels are unavailable. The update reward encourages confidence improvement and entropy reduction while penalizing excessive label drift, the retain reward provides positive feedback when updates would be harmful, and the abstain reward balances between avoiding premature updates and preventing policy collapse. This reward structure aligns well with the disambiguation objectives and provides meaningful learning signals.

  3. The method achieves competitive or superior results on both synthetic and real-world partial-label benchmarks. The reported improvements under increasing label ambiguity, together with component ablations and training-dynamics analyses, provide encouraging empirical support for the proposed framework.

Critical weaknesses
  1. Insufficient justification for reinforcement learning over simpler decision mechanisms. Although the Actor-Critic framework is presented as a central contribution, the paper does not sufficiently establish why sequential reinforcement learning is preferable to a deterministic reliability-based policy or a contextual bandit. The reported ablation does not fully isolate the benefits of long-term return optimization.

  2. Ambiguities in reward formulation and label-update consistency. The reward evaluates a proposal combining classifier predictions and neighborhood propagation, whereas the actual update applies only the neighborhood-propagated distribution. This discrepancy raises concerns about whether the learned policy is rewarded for the changes it actually makes.

  3. Limited reproducibility and evidence for reliability estimation. Several important implementation and theoretical details are deferred to appendices that are not included in the supplied manuscript. The reliability score is also used to determine curriculum admission and correction strength without sufficiently direct evidence that it reflects actual disambiguation correctness. While the overall results are promising, several gains over the strongest baselines are relatively modest. Considering the additional KNN propagation and Actor-Critic optimization, the computational cost relative to the achieved improvements remains insufficiently demonstrated.

Questions for authors
  1. (Most likely to change my score) Could the authors provide stronger evidence that the Actor-Critic policy offers benefits beyond simpler decision strategies, such as reliability-based heuristic selection or contextual bandits? In particular, a controlled comparison would help clarify whether long-term reward optimization is essential to the observed improvements.

  2. In Eqs. (8)–(14), the reward evaluates the fused proposal q^i\hat{q}_i, while the actual label update uses q~i\tilde{q}_i. Could the authors explain this discrepancy and clarify how the reward accurately reflects the quality of the executed correction?

  3. How well does the proposed reliability score correlate with actual label disambiguation correctness? Could the authors provide a quantitative analysis of its reliability, particularly for ambiguous or incorrectly predicted samples?

  4. Since both retain and abstain leave the current soft labels unchanged, what are their distinct roles in the sequential decision process? How do these actions affect subsequent learning differently?

  5. The curriculum threshold decreases linearly from 0.80 to 0.35. Could the authors explain the rationale for this schedule and whether the curriculum remains effective across datasets with different reliability distributions?

  6. Could the authors report training time and memory overhead compared with representative baselines? Additionally, do the authors have the intention of extending this to the PML field in the future?

2 comments
AuthorsOpenReview ↗

We thank the reviewer for the careful review and valuable feedback. We address the concerns below.

We thank the reviewer for the careful review and valuable feedback. We address the concerns below.

Q1: We conducted a controlled comparison on CIFAR-10 with q=0.3q=0.3. We compared five decision strategies:

  • Reliability-based heuristic: Update when ri≥max⁡(ηe,0.8)r_i\ge\max(\eta_e,0.8); retain admitted instances below the additional update threshold; and abstain on non-admitted instances.
  • Greedy reward selection: Select the legal action with the highest observable immediate reward, without learning a policy.
  • Contextual bandit: A learned stochastic policy optimized using immediate action rewards without temporal-difference bootstrapping.
  • Actor–Critic with γ=0\gamma=0: Retain the sampled-action Actor–Critic optimization while removing future-value bootstrapping.
  • Full Actor–Critic with γ=0.9\gamma=0.9: Include the value of the next observation, collected at the subsequent refinement round after classifier training, in the TD target.
Reliability-based heuristic85.90%
Greedy reward selection85.76%
Contextual bandit86.33%
Actor–Critic with γ=0\gamma=086.01%
Full Actor–Critic with γ=0.9\gamma=0.989.15%

The key difference between γ=0\gamma=0 and γ=0.9\gamma=0.9 is whether the policy foresees the delayed penalties of a correction. Experimental results show that the complete Actor–Critic strategy achieves the best disambiguation performance, we take this as direct evidence of the gains from optimizing long-term rewards.

Q2: The core motivation is that if the reward directly determined label updates, any error in the reward estimate would be written into the labels and self-reinforced in later rounds, causing cumulative error and error propagation. By decoupling reward computation from label updating, the reward only guides policy learning, while label updates are gated by the policy action and reliability constraints. The model can thus explore potential true labels in the action space while keeping actual updates conservative, preventing erroneous rewards from directly contaminating labels and reducing the risk of cumulative error.

Q3:

All instance

EpochCurrent soft-label accuracyKNN-corrected labelsReliability estimate corrected labels
5090.08%92.73%93.74%
10092.76%94.94%95.44%
15093.79%94.95%95.69%
20093.60%95.09%95.70%
25093.63%94.92%95.65%
30094.02%95.48%96.10%

High ambiguous instances

EpochCurrent soft-label accuracyReliability estimate corrected labels
5066.70%80.95%
10072.95%83.78%
15076.11%84.35%
20075.45%84.34%
25075.27%84.03%
30076.71%85.59%

We compare the current soft labels with both the KNN-corrected labels and the reliability estimate corrected labels, where the latter is obtained by applying the update operator to instance xi\mathbf{x}_{i}. The current-label accuracy increases from 90.08% at epoch 50 to 94.02% at epoch 300. Meanwhile, the disambiguation accuracy of both the KNN-corrected labels and the reliability estimate corrected labels improves significantly. These results demonstrate a strong empirical association between neighborhood reliability and label quality, and further indicate that reliability estimation can effectively improve disambiguation accuracy. We further evaluate the high-ambiguity instances. This result further demonstrates that reliability estimation consistently exhibits strong discriminative ability at the population level, effectively distinguishing the correctness of disambiguation results. This indicates that the update operator possesses an intrinsic potential for error correction. This reliability score consistently demonstrates a strong ability to discriminate the correctness of disambiguation at the population level, indicating that the update operator possesses an inherent potential for correction.

Q4:

Retain and abstain express different decision rationales but share the same direct label-state transition in our current implementation. Retain indicates that the proposed correction offers insufficient improvement, whereas abstain indicates insufficient evidence for correction. Their rewards reflect this distinction: the retain reward depends on the estimated proposal gain, while the abstain reward depends on reliability. Furthermore, the curriculum forces non-admitted instances to abstain. The original w/o-abstain experiment reports performance degradation. Our supplementary experiments on the CIFAR-10 dataset focus on comparing different decision-making strategies, rather than contrasting the two specific options within the "no-update" category of actions. The results show that both the full Actor-Critic method and a greedy selection strategy based on immediate rewards lead to improved final accuracy. This confirms that retaining both types of "no-update" actions is necessary, as it effectively prevents the accumulation of errors.

Q5:

We thank the reviewer for this question. The curriculum threshold gates correction by the reliability score: a high early threshold (0.80) restricts corrections to a small set of highly reliable samples to limit early error propagation, while the later lower threshold (0.35) expands coverage to moderately reliable samples once features and neighborhoods stabilize. The linear decrease provides a smooth risk transition and avoids sudden ILUR spikes; 0.80 and 0.35 were calibrated on validation data to correspond approximately to the high-reliability tail and the lower-middle reliability region, respectively. The effectiveness of this approach has been further verified through various data validation exercises and parameter sensitivity analyses.

Q6:

We sincerely thank the reviewer for this valuable suggestion. To assess computational efficiency, we conducted additional experiments comparing DC-RPLL with seven representative PLL methods in terms of training and inference GFLOPs and training time.

Training GFLOPs/sampleInference GFLOPs/sampleTime (min/epoch)
CC2.24400.74800.878
PRODEN2.24400.74801.087
LW2.24400.74801.087
CAVL0.21120.07040.264
VALEN0.45070.07040.403
ED-PLL2.24400.74801.178
MR-PLL2.24400.74801.618
DC-RPLL3.32901.11080.619

Although DC-RPLL requires approximately 48.4% more GFLOPs than CC, PRODEN, LW, ED-PLL, and MR-PLL, its measured training time is only 0.6192 min/epoch, representing reductions of 29.5%, 43.0%, and 61.7% compared with CC, PRODEN, and MR-PLL, respectively. This suggests that the additional computational complexity does not necessarily translate into proportionally higher training time. The peak GPU memory allocated by DC-RPLL, measured using PyTorch, is 832.93 MiB.

Regarding Partial Multi-Label Learning (PML), we consider it a promising future direction. Our reliability-guided selective correction could potentially be extended to instance–label-level decisions. However, accommodating multiple true labels requires redesigning the label representation, reliability estimation, and reward function to capture label dependencies. We therefore identify PML as future work rather than claiming direct applicability.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.