acceptodds
Under review as a conference paper at ICLR 2027

Breaking Majority Dominance: Progressive Abductive Learning under Concept Skewness

Abstract

Emulating the human interplay between perception and reasoning, abductive learning (ABL) is proposed to improve the perception model's ability in predicting concept labels without explicit annotations by drawing knowledge-consistent supervision signals from logical abduction, thereby benefiting reasoning. However, in practice, concept distributions are often skewed, leading to a biased feedback loop in ABL: early predictions favor majority concepts, making majority-dominated abduction candidates more likely to be reused as pseudo-labels to supervise the update of the perception model, further amplifying the bias against minority concepts. To solve this problem, we propose distribution-prior-rectified progressive abductive learning (DPRP-ABL). DPRP-ABL limits ambiguous candidate competition early in training by progressively expanding nested knowledge sub-bases. Within each active candidate set, it uses estimated concept distribution priors to alleviate majority-favoring scoring bias. Our analysis characterizes how these components affect candidate selection at a fixed model state and establishes conditions for preserving ground-truth feasibility across progressive phases. Experiments on two typical tasks indicate that DPRP-ABL improves reasoning-output performance and concept recognition under concept skewness.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Abductive learning, Concept skewness, Neuro-symbolic learning, Weakly supervised learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.60 over 5 reviews, above 70% of the others.
Reviewer_isArOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the Abductive Learning (ABL) framework that combines data-driven supervised machine learning with symbolic logical reasoning. Due to the skewed concept distributions (some correct but rare concepts may never be explored), the ABL process may get stuck in a vicious cycle, namely a bias feedback loop in ABL. This paper proposes distribution-prior-rectified progressive abductive learning (DPRP-ABL), which addresses this problem by: progression over the KB structure that can help slow down the bias feedback loop getting worse, and distribution-prior-award rectification that raises the weight of rare concepts to encourage the exploration of rare concepts. This paper presents a theoretical analysis to show that the effect of the candidate-size contraction, the effect of score rectification, and that the ground truth can always be covered. Experiments and ablation studies on digit addition and judicial sentencing demonstrate that DPRP-ABL is efficient and can outperform baselines.

Critical strengths

Contributions and significance

The studied Abductive Learning (ABL) framework is promising and important for combining data-driven supervised machine learning with symbolic logical reasoning. One of the main challenges of ABL is that due to the potentially noisy feedback from the symbolic reasoning, the ABL process may get stuck in a vicious cycle. This paper addresses this problem for imbalanced concept sets by progressively expanding the concept set and encouraging the exploration of rare concepts, which is of great significance for ABL.

About experiments

Extensive experiments are conducted on different datasets and different baselines. From the experiment results, it is shown that DPRP-ABL is efficient and outperforms baselines. The performance margins are significant for judicial sentencing.

Critical weaknesses
  • The difference of this paper from the related work should be presented more explicitly. The related work section should discuss the relationship and the differences between this paper and the related work rather than just listing the related work. In addition, the curriculum idea in ABL is proposed in Hu et al., 2025b, the authors should explicitly discuss the relationship and the differences.
  • The presentation of the grid search parameters should be clearly presented. There are many parameters in the experiments; the authors should clearly present the search space to enhance the reproducibility and the fairness of the experiments.
  • The scope of the theoretical results should be clearly presented. For example, the result about "Effect of score rectification" shows that rectification can change the ranking of the concepts, but this does not mean that rectification can improve the performance. The result of "Feasibility and Ground-truth Coverage Under Phase Transitions" shows that the ground truth can always be covered, but this does not mean the ground truth can always be found. I understand that the further analysis can be challenging, but these should be clearly discussed in the limitations section.

Other suggestions

  • The authors are recommended to add an example of the concept distribution skewness to let the readers know why this can be harmful and why it is important.
  • The authors are suggested to add more illustrations after the theoretical results to show the intuition and inspiration for the practice, rather than just showing the theorems.
Questions for authors

The distribution-prior-aware rectification is proposed to encourage the exploration of rare concepts. However, in some cases, this can lead to some incorrect concepts being explored.

  • How to evaluate whether the rectification is effective?
  • Can we implement a trade-off between exploration and exploitation?
5 comments
AuthorsOpenReview ↗

W1. Concerns about C-ABL and our DPRP-ABL.

Thanks for your suggestions. We agree on the importance of clarifying the relationship between our DPRP-ABL and C-ABL (Hu et al., 2025b). We acknowledge that the idea of the structure progressive component is inspired by the knowledge base curriculum introduced in C-ABL. However, we cannot ignore the differences between them in problem formulation, phase construction, transition criteria, and candidate scoring.

First of all, C-ABL introduces a knowledge base curriculum to reduce abduction space complexity and improve training stability. DPRP-ABL adopts this principle but studies a different failure mode, i.e., the majority-dominated biased feedback loop under concept skewness.

As for phase construction, for example, in the implementation of the digit addition task, our progression is organized by arithmetic structural complexity, i.e., by the number of carry propagations, which conforms to the laws of human learning that have been claimed at P1 in Section 3.2. You can refer to Appendix H.1 to know more about the metadata construction and progression design. By contrast, C-ABL organizes curriculum by activating the digit-concept domain. They offline construct metadata by manually partitioning concepts into gradually expandable subsets until the final stage covers all concepts. Consequently, one of the differences is that every phase in DPRP-ABL can retain all digits, while the admissible reasoning structures and candidate space are progressively expanded. In the implementation of the judicial sentencing task, our structural-complexity-based progression relies on the legal case complexity, based on which the metadata is constructed offline. The implementation proceeds by progressively expanding according to the implicated attributes within the cases until all cases are included in the sixth stage. This is consistent with C-ABL. Please refer to Appendix H.2 to obtain more details.

As for transition criteria among phases, for example, the C-ABL transition criterion uses held-out concept-level accuracy in the digit addition task and F1 in the judicial sentencing. This subset of concepts does not participate in the backpropagation of gradients during training as a supervisory signal. It is used solely for evaluation. By contrast, in the implementation on two tasks, DPRP-ABL advances according to agreement between current perception predictions and revised pseudo-labels in abduction, as you can see in Eq.(6). Therefore, DPRP-ABL does not require ground-truth concept labels during online training.

Last but not least, in terms of candidate scoring, DPRP-ABL adds a component of distribution-prior-aware rectification, which estimates the concept marginals and calculates weight to rectify candidate utilities to counter majority-dominated pseudo-label selection. Although C-ABL contracts the abduction space but does not explicitly model or handle challenges caused by skewed concept distribution.

To support these distinctions empirically, in addition to what already existed in Table 4 and Table 5 in the initial manuscript we submitted, we have also added experiments for an explicit comparison in Table 14 in Appendix I.1 for the digit addition task. Results suggest that our progression design may be particularly beneficial as perceptual difficulty and abduction ambiguity increase, while DPRP-ABL additionally rectify majority dominance in candidate competition, alleviating skewness-induced bias in the perception-abduction loop.

Besides, we will reorganize the narrative structure of the related work in the revision to clearly distinguish our approach from existing studies.


(Hu et al., 2025b) Wen-Chao Hu, Qi-Jie Li, Lin-Han Jia, Cun-Jing Ge, Yu-Feng Li, Yuan Jiang, and Zhi-Hua Zhou. Curriculum abductive learning, 2025b.

AuthorsOpenReview ↗

W2, W3, and S1

W2. Concerns about hyperparameter tuning.

Thanks for your suggestions. We agree on the necessity of providing details for the reproduction. We have added the details of hyperparameter tuning. Please refer to Fig. 15 in Appendix I.1.

In the digit addition task, we conduct the cross-grid search over β,τ∈{0.25,0.5,0.75,1}\beta,\tau\in\{0.25,0.5,0.75,1\}. Each grid cell reports the mean validation sum-level accuracy over the final five tuning epochs. The selected (β,τ)(\beta,\tau) values for d=1,2,3 are recorded in Fig. 15. In the judicial sentencing task, there is the same search range. (β,τ)=(1,1)(\beta,\tau)=(1,1) is selected according to the lowest mean absolute error (MSE).

For reproducibility and fairness, all methods share the same data splits, paired seeds, and training budgets. Experiments are conducted with the optimal hyperparameter that is tuned to ensure a fair comparison.

Besides, if you have more interest in reproducibility, please refer to the details in Appendix H.


W3. Concerns about the scope of score rectification and ground-truth coverage.

Thanks for your questions. We agree with both distinctions you raised.

Firstly, the score rectification result establishes only a condition under which the distribution prior term can change the ordering of two candidates. It does not establish that the rectified ranking agrees with the ground truth or that it necessarily improves downstream predictive performance. Rectification is beneficial when the original ordering is biased toward majority concepts and the estimated distribution prior provides a sufficiently accurate correction. When the distribution prior estimate or early perceptual predictions are noisy, it may over-correct the ranking and favor an incorrect minority candidate.

Secondly, feasibility or coverage is necessary but not sufficient for correct recovery. Even if the ground-truth assignment remains in the structural candidate set, another candidate may receive a higher combined perception-and-prior score and be selected. Thus, the coverage result does not imply that the ground truth will always be recovered.

We have moved the two corresponding remarks from the appendix to immediately after the score rectification and ground-truth coverage theorem, respectively. We now explicitly avoid interpreting a possible rank reversal as guaranteed performance improvement or structural coverage as guaranteed recovery.

Although these results do not constitute theoretical guarantees, the newly added analyses in Appendix I.1 provide direct empirical support for the intended mechanism. As shown in Table 13, Fig. 12, Fig. 13, and Fig. 14, minority pseudo-label recall increases together with minority precision and overall pseudo-label accuracy, while the minority proportion among selected candidates progressively approaches its ground-truth proportion. These results indicate that, in the evaluated settings, the proposed method improves the recovery of minority concepts without indiscriminately overpredicting them. The ablation, sensitivity, and majority-and-minority performance analyses provide further empirical evidence of its effectiveness.

Nevertheless, the present theory does not provide an end-to-end guarantee of correct pseudo-label selection, convergence, or generalization. Developing more reliable learning theoretic guarantees for the complete abductive learning process remains an important direction for our future work.


S1. Suggestions about adding an example of concept skewness.

Thanks for your suggestions. We agree that it would benefit from a more concrete example illustrating both the origin and the practical consequence of concept distribution skewness.

Fig. 1 provides an illustration of our research with a key perception-abduction feedback loop. In Fig. 4, the dashed curve shows the empirical marginals of the judicial attributes. As you can see, the distribution is highly skewed, where the positive rate of attitude is approximately 86.3%, whereas those of young, surrender, and forgive are only approximately 5.1%, 8.7%, and 11%, respectively. This skewness is harmful in abductive learning because an initially biased perception model tends to assign lower scores to candidates containing rare attributes. Consistency-based abduction may consequently select majority-dominated candidates as pseudo-labels, which provide even less subsequent supervision for the rare attributes and reinforce the initial bias. Importantly, low frequency does not imply low relevance. In judicial sentencing, relatively rare mitigating or aggravating circumstances may still materially affect the prediction, so outstanding aggregate performance can mask poor recognition of these attributes.

We will expand the discussion in the revision to make the motivation and practical importance of concept skewness more explicit.

AuthorsOpenReview ↗

Q1. Concerns about some incorrect concepts being explored and discussion about future work.

Thank you for raising this important issue. We agree that distribution prior rectification does not guarantee that every supported minority-containing candidate is correct. If the estimated distribution prior is inaccurate or the rectification strength is excessively large, an incorrect candidate containing rare concepts may receive an overly large score. We will clarify this limitation in the revision. Now, we would like to discuss with you our thoughts on the robustness of our DPRP-ABL regarding this issue, as well as suggestions for improving our future work.

We first emphasize that ''exploration'' in DPRP-ABL does not mean unconstrained generation or selection of rare concepts. Rectification only changes the relative scores of candidates that already satisfy the currently active knowledge constraints. Thus, a minority-containing candidate must remain logically feasible and sufficiently compatible with the perceptual predictions before it can be selected. The progressive knowledge base schedule provides an additional safeguard by restricting early abduction to a smaller and structurally simpler candidate space. We have added the evaluation to explain the effectiveness of rectification at the pseudo-label levels. As you can see in Table 13, Fig. 12, Fig. 13, and Fig. 14 in Appendix I.1, we report overall pseudo-label accuracy (ACC), minority pseudo-label recall (RECM) and precision (PREM) during training, the majority and minority composition of the candidate and selected sets, and the final performance of the majority concept and minority concept. Minority recall measures the benefit of recovering rare concepts, whereas minority precision directly measures the cost of incorrectly supporting them. The newly added training dynamics demonstrate that the increase in minority pseudo-label recall is accompanied by increasing minority precision and overall pseudo-label accuracy, rather than by indiscriminate overprediction of minority concepts. Moreover, as training progresses, the minority proportion among selected candidates approaches its ground-truth proportion, as you can see in Table 13 in Appendix I.1.

In this work, an exploration and exploitation trade-off can already be implemented through β\beta and τ\tau. Setting β=0\beta=0 or τ=0\tau=0 removes distribution-prior-based exploration and makes candidate selection rely primarily on perceptual likelihood, whereas larger values provide stronger compensation for rare concepts. We select these parameters using validation performance and have reported their sensitivity explicitly in Fig. 15 in Appendix I.1. Nevertheless, we agree that a more adaptive trade-off would be valuable. A promising extension is to use a phase-dependent or confidence-dependent coefficient βt\beta_t, increasing rectification only when the minority pseudo-label precision or abduction stability is sufficiently reliable and decreasing it when minority precision deteriorates. Another possibility is confidence-gated rectification, in which the rare concept bonus is applied only to candidates whose perceptual likelihood lies within a predefined margin of the best candidate. Clipping the maximum distribution-prior-rectification weight and maintaining estimates of the concept distribution prior could provide further safeguards.

We thank you again for your constructive suggestions and will discuss these adaptive strategies as an important direction for future work.

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

This paper studies abductive learning under concept distribution skewness and proposes DPRP-ABL. The method progressively expands knowledge sub-bases to limit early-stage candidate competition and applies inverse-prior correction to candidate scores using estimated concept marginals, thereby mitigating the amplification of majority-concept bias through pseudo-label feedback.

Critical strengths
  1. The motivation is clear and specific to ABL. Rather than treating the problem as conventional class imbalance in supervised learning, the paper explains how bias is amplified through the “perception prediction–knowledge-consistent candidate selection–pseudo-label training” pipeline and decomposes this process into competitive score dilution and supervision starvation.
  2. The experiments include relatively systematic ablation studies. Tables 3, 5, 8, and 9 evaluate different combinations of progressive scheduling and prior rectification. On the SVHN and judicial sentencing tasks, each component generally provides benefits when used independently, while the joint model achieves the best overall point estimates, suggesting that the two components are complementary to some extent.
  3. Some experimental improvements are substantial. DPRP-ABL achieves the best performance across all d=1,2,3d=1,2,3 settings on SVHN. On the judicial sentencing task, it also yields clear improvements on rare attributes such as theft and indoor.
Critical weaknesses
  1. It is unclear whether the phase construction uses hidden concept information.

    Section 3.2 constructs Dₜ using task-specific metadata aₙ and φ(aₙ, yₙ), but their exact definitions are not provided for either task. The natural complexity measure for the digit task is the number of carries, while that for the judicial task is the number of active attributes. If sample phases are determined using the ground-truth number of carries or active attributes, the procedure depends on hidden concepts and introduces additional supervision. Otherwise, the paper should provide an algorithm that relies only on X, y or candidate enumeration.

  2. The theoretical results are relatively weak.

    The theoretical phase sets are defined using complexity thresholds. However, the digit task additionally uses the minimum-Hamming-revision set in Eq. 20, while the judicial task further employs a revision budget, an optional textual compatibility check χ, and the dynamic Valid_ω constraint in Eq. 28. Each of these operations may exclude the ground truth. Consequently, the conditional ground-truth coverage result cannot be directly applied to the effective candidate sets used by Algorithm 1.

  3. The experiments do not directly validate the central minority-learning mechanism. The digit experiments report only aggregate concept-level metrics, without separate accuracy or recall results for the majority digits 0–4 and minority digits 5–9, or per-class results. The paper also does not report the candidate-set composition across iterations, overall pseudo-label accuracy, or minority-concept pseudo-label recall. Therefore, the current results demonstrate improvements in overall performance but do not directly verify the claimed mechanism of “breaking the majority-candidate feedback loop.”

  4. The reproducibility and statistical information are insufficient.

    The paper does not fully specify the train/validation/test sample counts and splits, the sampling procedure for the labeled subset in the judicial task, or the implementation details and search ranges for β, τ, T, bₜ, and the phase-advancement threshold. The main results report only means and standard deviations over five runs, without paired confidence intervals or significance tests. Moreover, the variance is relatively large on CIFAR and for some judicial attributes.

Questions for authors
  1. Please provide the exact construction algorithms for aₙ, φ(aₙ, yₙ), Dₜ, and KBₜ in both tasks. At any stage, do you use ground-truth concept labels or any function of those labels to determine the number of carries in the digit task or the number of active attributes in the judicial task?

  2. What is the algorithmic difference between S1 and C-ABL? Could the authors compare C-ABL, C-ABL with post-hoc inverse-prior logit adjustment, and DPRP-ABL using the same backbone, phase metadata, and training budget?

  3. Please directly validate the feedback-loop mechanism proposed in the paper. Could the authors report minority pseudo-label recall and precision at each iteration, overall pseudo-label accuracy, the composition of majority and minority concepts in the candidate sets, and comparative training dynamics on balanced and skewed datasets?

  4. Please provide grouped and per-class results for the digit task, particularly accuracy, recall, and macro-F1 for digits 0–4 versus 5–9. A balanced test set could also be included as a diagnostic experiment to determine whether the improvements indeed arise from better minority-concept learning.

  5. Please provide complete details of the data splits and hyperparameter tuning, including the train/validation/test sample counts for both tasks, the sampling strategy for the 10% and 50% labeled subsets in the judicial task, whether paired random seeds were used across the five runs, and the search ranges for all key hyperparameters.

9 comments
AuthorsOpenReview ↗

Q1. Concerns about metadata construction and progression design.

Thanks for your questions. We clarify that the metadata are constructed entirely offline and are fixed before DPRP-ABL training begins. This general idea is inspired by the curriculum construction of C-ABL (Hu et al., 2025b). However, we adapt this idea to the structural characteristics of the two tasks according to our proposed principles P1-P3, which are detailed in Section 3.2. These principles organize the data and knowledge from simple to complex, making the progression compatible with the data-and-knowledge-driven learning paradigm of abductive learning and with an easy-to-hard learning process. The metadata are generated through an offline expert or domain-knowledge-engineering process, rather than by querying ground-truth concepts during online learning.

Specifically, for the digit-addition task, we use independently trained and frozen expert classifiers to recognize the input images and construct the carry-complexity metadata. The ground-truth digit concepts of the DPRP-ABL training samples are never queried by the progression constructor. Based on the expert-generated metadata, equations are assigned to the no-carry phase, the at-most-one-carry phase, and the unrestricted final phase, which restores the complete training set and knowledge base. For the judicial sentencing task, the structural metadata are provided as part of the domain-expert knowledge-engineering process, where legal experts organize previously adjudicated cases according to their structural complexity, based on the number and configuration of relevant legal factors. Cases are assigned to increasingly complex phases, with the final phase containing all cases. In both tasks, the online learner receives only a phase identifier or eligibility mask, rather than the concept labels used to describe the sample. These metadata determine when a sample becomes active and which knowledge sub-base is used, but they are never used as concept-level supervision or as additional numerical terms in the perception loss, candidate scoring, or online phase transition decision.

Once the offline metadata have been fixed, phase advancement is determined automatically from the online learning state of the model. In the digit addition task, we measure abduction stability as the agreement between the perception predictions and the revised pseudo-labels. DPRP-ABL advances when the mean stability reaches the threshold or the maximum per-phase training budget. In the judicial sentencing task, advancement is also conducted when the stability is met or the training budget is reached. Neither transition criterion uses ground-truth concepts.

We will expand Appendix H by adding details about dataset splitting, expert models, metadata construction, and progression design for both tasks.


(Hu et al., 2025b) Wen-Chao Hu, Qi-Jie Li, Lin-Han Jia, Cun-Jing Ge, Yu-Feng Li, Yuan Jiang, and Zhi-Hua Zhou. Curriculum abductive learning, 2025b.

AuthorsOpenReview ↗

Q2. Concerns about C-ABL and in our DPRP-ABL.

Thanks for your questions. There are two main differences between S1S_1 and C-ABL.

Firstly, that is phase construction. For example, in the implementation on the digit addition task, S1S_1 uses a carry-structured progression, i.e., the number of carry propagations, which conforms to the laws of human learning that have been claimed at P1 in Section 3.2. You can refer to Appendix H.1 to know more about the metadata construction and progression design. By contrast, C-ABL organizes curriculum by activating the digit-concept domain. They offline construct metadata by manually partitioning concepts into gradually expandable subsets until the final stage covers all concepts. Consequently, one of the differences is that every phase in S1S_1 can retain all digits, while the admissible reasoning structures and candidate space are progressively expanded. In the implementation of the judicial sentencing task, S1S_1 uses a legal case-structured progression. The implementation proceeds by progressively expanding according to the implicated attributes within the cases until all cases are included in the sixth stage. This is consistent with C-ABL. You can refer to Appendix H.2 to obtain more details.

However, what we cannot ignore is the second difference. That is the transition criterion among phases. The C-ABL transition criterion uses held-out concept-level accuracy in the digit addition task and F1 in the judicial sentencing. By contrast, in the implementation on two tasks, S1S_1 advances according to agreement between current perception predictions and revised pseudo-labels in abduction, as you can see in Eq. (6). Therefore, S1S_1 does not require ground-truth concept labels during online training.

To empirically clarify these distinctions, we complement the judicial-task comparisons already reported in Table 4 and Table 5 with explicit comparisons on the digit-addition task. Table 14 in Appendix I.1 reports comparisons among C-ABL, DPRP-ABL with only S1S_1, and DPRP-ABL across three perception datasets with d∈{1,2,3}d\in\{1,2,3\}. Table 15 in Appendix I.1 further provides a controlled comparison on the SVHN dataset among C-ABL, C-ABL with inference-only post-hoc inverse-prior adjustment, DPRP-ABL with only S1S_1, and DPRP-ABL. All methods use the same backbone, data splits, paired random seeds, number of progression phases, and total training budget. Their curriculum metadata nevertheless follows the respective method designs, i.e., C-ABL progressively expands the active concept domain, whereas DPRP-ABL organizes its phases according to carry complexity.

As for results reported in Table 14 in Appendix I.1, comparing C-ABL with the S1S_1-only variant, the latter achieves higher z_ACC, PREM, and y_ACC across d=1,2,3 on the SVHN and CIFAR datasets in most cases, suggesting that our overall design, i.e., combining carry complexity-based metadata with stability-based phase advancement, can be beneficial under challenging perception and substantial abduction ambiguity. Comparing the S1S_1-only variant with DPRP-ABL, which share the same metadata and progression rule, their difference isolates the contribution of distribution-prior-aware rectification. As for results reported in Table 15 in Appendix I.1, unlike inference-only post-hoc adjustment, which modifies only the final predictions of a trained model, DPRP-ABL incorporates the estimated concept distribution prior into candidate scoring in abduction and pseudo-label generation throughout training. It consistently outperforms the post-hoc baseline for all three sequence lengths. These results indicate that post-hoc adjustment alone does not reproduce the broader gains of DPRP-ABL and support addressing concept skewness directly within the perception-abduction loop.

AuthorsOpenReview ↗

Q3. Concerns about the feedback-loop mechanism in our DPRP-ABL.

Thanks for your suggestions. We agree that it is important to establish how candidate competition and pseudo-label supervision evolve during training as direct process-level evidence for the feedback-loop mechanism. We have added epoch-wise pseudo-label diagnostics, phase-wise candidate set statistics, and comparative training dynamics under balanced and skewed concept distributions. Please refer to Table 13, Fig. 12, Fig. 13, and Fig. 14 in Appendix I.1.

We first tell you our definition of ''MinRR'' as the per-concept sample-count ratio of minority to majority digits. Values of 0.05, 0.1, 0.5, and 1 correspond to majority-to-minority ratios of 20:1,10:1,2:120{:}1, 10{:}1, 2{:}1, and 1:11{:}1, respectively. A larger value indicates weaker skewness. There is MinRR=0.05\text{MinRR}=0.05 in the initial manuscript we submitted.

As for candidate space contraction, please refer to Table 13. Under the most severe skewness setting, i.e., MinRR=0.05\text{MinRR}=0.05, the mean candidate set sizes before the first update are 2.32, 20.14, and 90.29 for d=1,2,3, respectively, confirming that abduction ambiguity increases rapidly with the length of the reasoning chain. During the first phase, these values decrease to 1.41, 3.99, and 5.32, and they further approach a single candidate in the final phase (i.e., 1.00, 1.00, and 1.02). The contraction is particularly substantial for d=2 and d=3, where the initial candidate set sizes are reduced by approximately 95% and 98.9% by the final phase.

As for the minority coverage during early candidate competition, please refer to Table 13. For d=2, minority concepts constitute only 2.96% of the active ground-truth assignments in Phase 1 but account for 14.60% of the feasible candidate occurrences and 13.06% of the selected pseudo-label occurrences. Similarly, for d=3, the corresponding ground-truth, candidate, and selected minority ratios are 2.86%, 13.59%, and 13.28%. Thus, the early candidate pool does not simply reproduce the approximately 97% majority composition of the active data. Instead, it preserves substantially greater coverage of minority-containing alternatives, preventing them from being excluded from candidate competition. We emphasize that candidate composition measures coverage rather than correctness. Correctness is evaluated separately by the pseudo-label precision and recall in our following analyses.

As for pseudo-label quality, Fig. 12, Fig. 13, and Fig. 14 report overall position-wise pseudo-label accuracy (ACC) and its minority-concept precision (PREM) and recall (REC) at each epoch. Across all sequence lengths and minority-to-majority retention ratio (MinRR) settings, pseudo-label accuracy and minority recall increase together during training. Minority precision also improves, indicating that the recall gain is not achieved by persistently over-predicting minority concepts. This progressive improvement in minority pseudo-label recall and precision is consistent with the proposed mitigation of minority supervision starvation. Under more severe skewness, minority recall can initially exceed minority precision, which is consistent with the temporary minority enrichment observed in the candidate sets. This gap subsequently narrows as the model improves. Importantly, the selected pseudo-label composition becomes closely aligned with the ground-truth composition in the final phase. The selected versus ground-truth minority ratios are ''4.69% versus 4.66%'', ''4.89% versus 4.87%'', and ''4.90% versus 4.79%'' for d=1,2,3, respectively. Therefore, the enhanced early coverage of minority concepts does not result in persistent minority over-prediction. Instead, it is gradually converted into accurate pseudo-supervision whose concept composition closely matches the ground truth distribution.

As for balanced versus skewed dynamics, in Fig. 12, Fig. 13, and Fig. 14, we compare MinRR∈{0.05,0.1,0.5,1}\text{MinRR}\in\{0.05,0.1,0.5,1\}, where MinRR=1\text{MinRR}=1 is concept-balanced. In every setting, the three pseudo-label quality metrics improve and stabilize during training. Severely skewed settings generally exhibit a larger early recall-precision gap and greater variability, whereas the balanced setting shows more closely aligned minority recall and precision. Nevertheless, DPRP-ABL does not exhibit a persistent minority-recall collapse even at MinRR=0.05\text{MinRR}=0.05, indicating that our method remains effective under the skewed setting where the majority-dominated feedback loop is expected to be heaviest.

To sum up, these results provide process-level evidence for the proposed mechanism. Candidate space contraction reduces early abduction ambiguity. Minority-containing alternatives remain available during candidate competition, which are converted into increasingly accurate minority pseudo-labels. The selected pseudo-label distribution is ultimately calibrated to the active ground-truth concept distribution.

Reviewer_D8XnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes distribution-prior-rectified progressive abductive learning (DPRP-ABL) to alleviate the concept skewness issue in abductive learning, which is formulated by the authors. Specifically, early predictions favor majority concepts, which causes the phenomenon that the majority-dominated abduction candidates are more likely to be used later in the updates of the perception model. DPRP-ABL narrows the number of candidates by using structure-progressive scheduling, limiting early abduction to nested knowledge sub-bases, and combines structure-progressive knowledge scheduling with distribution-prior-aware rectification. Experimental results justify the outstanding performance of DPRP-ABL.

Critical strengths

None

Critical weaknesses

I would be open-minded if the authors could clarify the following points.

  • C-ABL, a strong baseline that seems very relevant to DPRP-ABL and uses progressive ABL, is missing from Table 1. Why?
  • Line 181: "Denote ana_n as task-specific structural metadata available to the offline progression constructor." For the training data, where does this metadata come from?
  • Line 307-309: Such concept skewness simulation seems specific to the digit addition task and heuristic. Can you further explain the reasons behind this setup, or run experiments under more skewness setups to show the generalization ability of DPRP-ABL?
Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Q1. Concerns about C-ABL and our DPRP-ABL.

Thanks for your questions. We agree that C-ABL is a highly relevant baseline and should have been discussed more prominently. In Table 1, we do not consider C-ABL as a comparison because of an evaluation-protocol ambiguity rather than a lack of relevance. That is about transition criteria among phases. The published C-ABL (Hu et al., 2025b) phase-transition rule monitors ground-truth concept accuracy on a held-out validation set, whereas our setting assumes that digit concepts remain unavailable throughout online training. Although concept labels in the held-out validation set are not involved in training as supervision signals contributing to gradient propagation, and they are used solely for evaluation, we think it might be unfair. We advocate that during the task processing, this information should remain entirely invisible. It does not affect the gradients in model training, nor should it influence the online training progression, and ground-truth concept labels should be accessible only offline.

To address this concern without introducing additional concept supervision during training processing, we have implemented a fixed-epoch transition adaptation of C-ABL using its concept-domain curriculum. This way is acknowledged by their authors. You can find details in their rebuttal on OpenReview. Results in the digit addition task are reported in Table 14 in Appendix I.1. Besides, for the judicial sentencing task, existing ABL methods capable of adapting to this task are scarce, so even if we may be disadvantaged due to unfair factors, we nonetheless proceed to compare. However, a fair comparison can be judged by reference to the ablation experiments in Table 5 where only S1S_1 is present. In this task, our metadata and nested sub-bases are the same. The only difference between S1S_1 and C-ABL is that the transition criteria of S1S_1 are based on the agreement between current perception predictions and revised pseudo-labels in abduction, which is proposed in Eq. (6). It can, to a certain extent, substitute for C-ABL in comparison with DPRP-ABL.

Results demonstrate that our progression design may be particularly beneficial as perceptual difficulty and abduction ambiguity increase, while DPRP-ABL additionally rectify majority dominance in candidate competition, alleviating skewness-induced bias in the perception-abduction loop.

Table 14. Comparison ''Avg±std'' with C-ABL on the digit addition task.

MethodsdMNIST z_ACC (↑)MNIST AUC (↑)MNIST PREM (↑)MNIST y_ACC (↑)SVHN z_ACC (↑)SVHN AUC (↑)SVHN PREM (↑)SVHN y_ACC (↑)CIFAR z_ACC (↑)CIFAR AUC (↑)CIFAR PREM (↑)CIFAR y_ACC (↑)
C-ABL10.9664±0.00560.9972±0.00050.9687±0.00480.9353±0.01060.8600±0.00740.9776±0.00160.8741±0.00490.7412±0.01280.3752±0.01020.7423±0.00750.5400±0.01810.1612±0.0069
DPRP-ABL (only S₁)10.9664±0.00930.9970±0.00160.9689±0.00830.9357±0.01740.9014±0.00590.9842±0.00090.9054±0.00460.8130±0.01020.3344±0.09770.7213±0.07130.4191±0.17320.1344±0.0648
DPRP-ABL10.9721±0.00160.9987±0.00030.9736±0.00120.9461±0.00270.9023±0.00020.9846±0.00050.9056±0.00080.8149±0.00050.3817±0.06190.7622±0.03910.5539±0.08040.1588±0.0450
C-ABL20.9789±0.00120.9985±0.00040.9797±0.00120.9199±0.00450.9024±0.00400.9874±0.00060.9069±0.00380.6643±0.01150.4825±0.01650.8171±0.00810.6093±0.01990.0616±0.0069
DPRP-ABL (only S₁)20.9780±0.00300.9987±0.00080.9790±0.00300.9174±0.01110.9176±0.00260.9875±0.00100.9187±0.00210.7099±0.00860.5943±0.08440.8706±0.03410.6898±0.04340.1403±0.0692
DPRP-ABL20.9789±0.00280.9990±0.00060.9799±0.00250.9206±0.01010.9262±0.00460.9889±0.00100.9270±0.00450.7360±0.01400.6176±0.03160.8865±0.01610.7068±0.01730.1515±0.0292
C-ABL30.9812±0.00170.9990±0.00020.9818±0.00170.8945±0.00940.9200±0.00480.9905±0.00080.9213±0.00460.6084±0.01870.5382±0.07580.8512±0.04160.6263±0.06150.0343±0.0196
DPRP-ABL (only S₁)30.9809±0.00330.9986±0.00060.9815±0.00330.8927±0.01770.9244±0.00090.9859±0.00080.9250±0.00090.6247±0.00280.6706±0.08320.9033±0.03090.7155±0.05500.1099±0.0483
DPRP-ABL30.9809±0.00150.9994±0.00020.9819±0.00120.8934±0.00780.9297±0.00180.9890±0.00090.9299±0.00190.6468±0.00750.7238±0.00950.9251±0.00840.7469±0.00200.1497±0.0117

(Hu et al., 2025b) Wen-Chao Hu, Qi-Jie Li, Lin-Han Jia, Cun-Jing Ge, Yu-Feng Li, Yuan Jiang, and Zhi-Hua Zhou. Curriculum abductive learning, 2025b.

AuthorsOpenReview ↗

Q2. Concerns about metadata.

Thanks for your questions. We clarify that the metadata are constructed entirely offline and are fixed before DPRP-ABL training begins. This general idea is inspired by the curriculum construction of C-ABL (Hu et al., 2025b). However, we adapt this idea to the structural characteristics of the two tasks according to our proposed principles P1-P3, which are detailed in Section 3.2. These principles organize the data and knowledge from simple to complex, making the progression compatible with the data-and-knowledge-driven learning paradigm of abductive learning and with an easy-to-hard learning process. The metadata are generated through an offline expert or domain-knowledge-engineering process, rather than by querying ground-truth concepts during online learning.

Specifically, for the digit-addition task, we use independently trained and frozen expert classifiers to recognize the input images and construct the carry-complexity metadata. The ground-truth digit concepts of the DPRP-ABL training samples are never queried by the progression constructor. Based on the expert-generated metadata, equations are assigned to the no-carry phase, the at-most-one-carry phase, and the unrestricted final phase, which restores the complete training set and knowledge base. For the judicial sentencing task, the structural metadata are provided as part of the domain-expert knowledge-engineering process, where legal experts organize previously adjudicated cases according to their structural complexity, based on the number and configuration of relevant legal factors. Cases are assigned to increasingly complex phases, with the final phase containing all cases. In both tasks, the online learner receives only a phase identifier or eligibility mask, rather than the concept labels used to describe the sample. These metadata determine when a sample becomes active and which knowledge sub-base is used, but they are never used as concept-level supervision or as additional numerical terms in the perception loss, candidate scoring, or online phase transition decision.

We will expand Appendix H by adding details about dataset splitting, expert models, metadata construction, and progression design for both tasks.


(Hu et al., 2025b) Wen-Chao Hu, Qi-Jie Li, Lin-Han Jia, Cun-Jing Ge, Yu-Feng Li, Yuan Jiang, and Zhi-Hua Zhou. Curriculum abductive learning, 2025b.

AuthorsOpenReview ↗

Q3. Concerns about concept skewness settings.

Thanks for your suggestions. Let us explain the rationale for our setting of simulated concept skewness. Furthermore, we have added evaluation under more skewness levels to show the generalization ability of DPRP-ABL. Please refer to Table 12 in Appendix I.1.

We selected digit addition because it is a widely used controlled benchmark for weakly supervised neuro-symbolic learning and abductive learning. The task contains raw perceptual inputs, latent digit concepts, explicit symbolic addition rules, and supervision available only at the sum level. It directly reflects the data-and-knowledge-driven motivation of abductive learning, i.e., the perception model must recognize unobserved concepts from images, while logical abduction uses the arithmetic rules and observed sums to revise concept predictions. Related tasks have been used in representative neuro-symbolic and abductive learning studies (Manhaeve et al., 2018; Dai et al., 2019). Unlike the judicial sentencing task, where attribute skewness arises naturally from real case frequencies, the original digit datasets are approximately class-balanced. A controlled sampling procedure is therefore necessary to study abductive learning under concept skewness and to vary skewness independently of the perceptual and logical components. Specifically, our construction follows the commonly used step-imbalance protocol, i.e., half of the concepts (i.e., digits ''0''-''4'') form the majority group, while the remaining half (i.e., digits ''5''-''9'') form the minority group. Concepts within each group have the same sampling frequency. This construction, including the use of an equal fraction of majority and minority classes, has been adopted in controlled imbalanced-learning studies because it provides a clear distinction between two groups and avoids additional within-group frequency heterogeneity (Cao et al., 2019). The original minority-to-majority retention ratio (MinRR) of 0.05 corresponds to a majority-to-minority imbalance factor of 20. We intended this as a pronounced controlled stress test rather than as a claim that digit frequencies naturally follow this distribution.

We agree that more experiments simulating different degrees of skewness would be helpful in verifying the generalization ability of DPRP-ABL. Following your suggestion, we additionally evaluate DPRP-ABL using minority-to-majority retention ratios (MinRR) of 0.05, 0.1, and 0.5, corresponding to imbalance factors of 20, 10, and 2, respectively. These settings cover progressively weaker levels of concept skewness. DPRP-ABL remains effective across all three skewness levels. Even under the most severe imbalance factor of 20, the concept-level accuracy (z_ACC) ranges from 0.9023 to 0.9297, AUC ranges from 0.9846 to 0.9890, and G-mean ranges from 0.9443 to 0.9604 across d=1,2,3. Thus, the model does not collapse when only 5% as many minority concept samples as majority concept samples are retained. As levels of concept skewness are weaker, concept-level accuracy improves by 2.60%, 1.53%, and 0.96% for d=1,2,3, respectively. Target-level accuracy (y_ACC) improves by 4.73%, 5.03%, and 4.19%. Although a few ranking-based metrics exhibit small non-monotonic fluctuations, the overall trend is consistent, i.e., weaker skewness improves performance, while DPRP-ABL degrades gracefully as concept skewness becomes more severe. These results show that the effectiveness of DPRP-ABL is not restricted to the single MinRR=0.05 sampling configuration.

Besides, if you have interest, please refer to Fig. 12, Fig. 13, and Fig. 14 in Appendix I.1, where additional metric trajectories under the three skewness levels are reported, providing a complementary view of training dynamics in the perception-abduction loop.


(Manhaeve et al., 2018) Robin Manhaeve, Sebastijan Dumancic, Angelika Kimmig, Thomas Demeester, and Luc De Raedt. DeepProbLog: Neural probabilistic logic programming, 2018.

(Dai et al., 2019) Wang-Zhou Dai, Qiu-Ling Xu, Yang Yu, and Zhi-Hua Zhou. Bridging machine learning and logical reasoning by abductive learning. In Proceedings of the Annual Conference on Neural Information Processing Systems, volume 32, pp. 2811-2822, 2019.

(Cao et al., 2019) Kai-Di Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Teng-Yu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In Proceedings of the Annual Conference on Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019.

AuthorsOpenReview ↗

Q3. Concerns about concept skewness settings. Experiment results in Table 12.

Table 12. Robustness of DPRP-ABL under different levels of concept distribution skewness on the SVHN dataset in the digit addition task. MinRR is the minority-to-majority retention ratio, and a larger value indicates weaker skewness.

dMinRRz_ACC (↑)AUC (↑)RECM (↑)PREM (↑)MAP (↑)G-mean (↑)y_ACC (↑)
10.050.9023±0.00020.9846±0.00050.9023±0.00020.9056±0.00080.9368±0.00140.9443±0.00010.8149±0.0005
10.10.9159±0.00520.9874±0.00070.9159±0.00520.9171±0.00490.9463±0.00810.9523±0.00310.8390±0.0093
10.50.9283±0.00680.9891±0.00140.9283±0.00680.9287±0.00670.9508±0.00570.9596±0.00390.8622±0.0126
20.050.9262±0.00460.9889±0.00100.9262±0.00460.9270±0.00450.9577±0.00400.9583±0.00270.7360±0.0140
20.10.9295±0.00260.9885±0.00080.9295±0.00260.9300±0.00240.9563±0.00360.9603±0.00150.7490±0.0075
20.50.9415±0.00250.9918±0.00130.9415±0.00250.9416±0.00260.9623±0.00700.9671±0.00140.7863±0.0081
30.050.9297±0.00180.9890±0.00090.9297±0.00180.9299±0.00190.9607±0.00260.9604±0.00100.6468±0.0075
30.10.9317±0.00740.9886±0.00120.9317±0.00740.9322±0.00750.9623±0.00080.9615±0.00420.6545±0.0301
30.50.9393±0.00310.9919±0.00090.9393±0.00310.9395±0.00320.9620±0.00440.9658±0.00180.6887±0.0146
Reviewer_rMMvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Abductive Learning (ABL) struggles when latent concepts are heavily skewed: early predictions favor majority concepts, so majority-dominated candidates win the abduction step ("competitive score dilution") and minority concepts receive no pseudo-label supervision ("supervision starvation"), which reinforces the bias. The authors propose DPRP-ABL, which trains progressively over nested knowledge sub-bases from simple to complex rules and rectifies candidate scores with inverse-prior weights based on estimated concept marginals, obtained either from the target distribution or from a partially labeled subset. The authors support the method with three fixed-state theoretical propositions and evaluate it on skewed d-digit addition (MNIST, SVHN, CIFAR) and a semi-supervised judicial sentencing task.

Critical strengths
  • Well-Motivated Problem Formulation: The paper clearly articulates why class imbalance in ABL differs from standard supervised imbalance. Because supervision is mediated by combinatorial candidate competition over S={z∈Zm∣z∧KB⊨y}\mathcal{S} = \{z \in \mathcal{Z}^m \mid z \land \mathcal{KB} \models y\}, bias scales multiplicatively across sequence length mm (Eq. 3), rather than merely affecting isolated per-instance decision boundaries.

  • Complementary Dual-Axis Design: The coupling of abduction-space contraction (Axis 1) and prior-based score rectification (Axis 2) is intuitive. The 2×22 \times 2 ablation studies across both tasks (Tables 3, 5, 8, and 9) consistently demonstrate that combining S1S_1 and S2S_2 outperforms applying either mechanism in isolation.

  • Versatility Across KB Modalities: The framework is successfully instantiated on both a discrete, deterministic symbolic KB (dd-digit addition) and a parametric continuous KB co-optimized with perception parameters (judicial sentencing).

  • Strong Empirical Performance and Efficiency: DPRP-ABL achieves strong target-level and concept-level gains, particularly as perceptual difficulty and reasoning horizon increase (e.g., all SVHN settings and multi-digit CIFAR with d∈{2,3}d \in \{2,3\} in Table 1). Notably, it occasionally surpasses the unrectified fully supervised cross-entropy baseline (SUP) in concept AUC (Table 2), while substantially reducing per-epoch training time (Table 1) and token consumption to convergence (Table 4).

Critical weaknesses
  1. Potential Ground-Truth Concept Leakage in Sample Stratification (Major Concern):

    In Section 3.2 and Algorithm 1, the active dataset at phase t is defined as Dₜ = {(Xₙ, yₙ) : φ(aₙ, yₙ) ≤ t}, where aₙ is described as "task-specific structural metadata available to the offline progression constructor." However, Appendices E.1 and E.2 define the complexity measures exclusively over the latent concept vector z: κ_add(z) = carry(z) for addition (Eq. 17) and κ_jud(z) = ‖z‖₀ for judicial sentencing (Eq. 26). Furthermore, Theorem 3 requires κ(z*) ≤ bₜ to guarantee that the ground-truth assignment z* remains feasible in Aₜ ⊆ Sₜ.

    This raises a fundamental concern about the availability of the information used for sample stratification:

    • Multi-digit addition: The number of carries cannot, in general, be uniquely determined from the observed sum yₙ alone. For example, y = 10 can arise from 5 + 5, which involves one carry, or from 10 + 0, which involves no carry.

    • Judicial sentencing: Similarly, ‖z*‖₀ is a direct function of the ground-truth attributes, which are assumed to be unannotated and unavailable during training.

    Consequently, if Dₜ filters samples according to κ(zₙ*) ≤ bₜ, then aₙ must encode information about the unobserved ground-truth concepts zₙ*. Although aₙ is not directly provided to f_θ, using it to partition Dₜ and match the resulting subsets with KBₜ based on κ(zₙ*) effectively introduces privileged ground-truth supervision that is not available to the baselines.

    Conversely, if Dₜ = D throughout all phases and only the candidate space Sₜ(y) is restricted by κ(z) ≤ bₜ, then samples satisfying κ(zₙ*) > bₜ will have their ground-truth explanations excluded from the candidate space during early phases. This potentially forces the model to produce incorrect pseudo-labels and appears inconsistent with the feasibility condition required by Theorem 3.

    Importantly, Appendix E does not explicitly define aₙ or φ(aₙ, yₙ), leaving this critical aspect of the proposed training procedure ambiguous. The authors should clarify exactly how these quantities are constructed for both tasks and demonstrate that no ground-truth concept information is used in sample stratification.

  2. Limited Depth of Theoretical Contributions:

    The three theorems presented in Section 4 appear to provide relatively elementary mathematical properties rather than substantial learning-theoretic guarantees. In particular:

    • Theorem 1: This result essentially establishes that removing non-negative terms from the denominator of a softmax distribution (or restricting a finite candidate set around its maximizer) weakly increases the probability assigned to the remaining elements. While mathematically valid, this is a relatively straightforward consequence of probability normalization and does not, by itself, provide a strong theoretical guarantee about the learning dynamics or generalization behavior of the proposed method.
Questions for authors
  1. Clarification on φ(aₙ, yₙ) and Potential Leakage: How are aₙ and φ(aₙ, yₙ) explicitly defined and computed in Tasks 1 and 2? Does constructing Dₜ use ground-truth carry counts, carry(zₙ*), or ground-truth attribute sparsity, ‖zₙ*‖₀, to filter samples per phase? If so, how do you justify using latent ground-truth concept properties to partition the training set when concept labels are assumed unavailable? If not (i.e., if Dₜ = D at every phase and only Sₜ is constrained), what happens to samples where κ(zₙ*) > bₜ during early phases when Sₜ(yₙ) is either empty or excludes zₙ*?

  2. Clarification on the C-ABL Baseline and Generalizability: Is the "S₁ only" (S₁ = ✓, S₂ = ✗) ablation in Tables 3, 8, and 9 identical to C-ABL (Hu et al., 2025b) on the d-digit addition task? If there are algorithmic differences, why was C-ABL excluded from Table 1? Additionally, can ImbNSL be adapted to the judicial sentencing task?

  3. Prior Estimation and Sensitivity Analysis: In Task 1, the ground-truth digit prior is πᵏ ≈ 0.1905 for k ∈ {0, ..., 4} and πᵏ ≈ 0.0095 for k ∈ {5, ..., 9}. What are the actual estimated values r̂ recovered by solving Eq. (7) for d = 1, 2, 3? How sensitive is DPRP-ABL to estimation errors in r̂, especially in Task 2 at 10% labeled data, where early pseudo-labels Pₑ in Eq. (29) may be noisy?

  4. Hyperparameter Specification and Sensitivity Analysis: Given that β log(vᵏ) = −βτ log(r̂ᵏ) − β log(Zᵥ), why are both β and τ retained as separate hyperparameters? Please provide the exact values of β, τ, phase complexity thresholds {bₜ}ₜ₌₁ᵀ, and phase advancement criteria ηₜ used for each dataset, along with a sensitivity analysis for β and τ.

6 comments
AuthorsOpenReview ↗

Q1. Concerns about metadata construction.

Thank you for raising this important question. We clarify that the metadata are constructed entirely offline and are fixed before DPRP-ABL training begins. This general idea is inspired by the curriculum construction of C-ABL (Hu et al., 2025b). However, we adapt this idea to the structural characteristics of the two tasks according to our proposed principles P1-P3, which are detailed in Section 3.2. These principles organize the data and knowledge from simple to complex, making the progression compatible with the data-and-knowledge-driven learning paradigm of abductive learning and with an easy-to-hard learning process. The metadata are generated through an offline expert or domain-knowledge-engineering process, rather than by querying ground-truth concepts during online learning.

Specifically, in the digit addition task, ana_n consists of the digit predictions produced by independently trained and frozen visual experts, and ϕ(an,yn)=κ(z~n)\phi(a_n,y_n)=\kappa(\widetilde{\mathbf z}_n), where z~n\widetilde{\mathbf z}_n is the expert-supported digit assignment and κ\kappa counts carry operations. More precisely, z~n∈S(yn)\widetilde{\mathbf z}_n\in S(y_n) is selected from the assignments consistent with the observed sum yny_n, using the frozen expert predictions. Therefore, the offline phase constructor does not query carry⁡(zn∗)\operatorname{carry}(\mathbf z_n^*). Based on such carry complexity, equations are assigned to the no-carry phase, the at-most-one-carry phase, and the unrestricted final phase. Formally, letting κ^n=κ(z~n)\widehat{\kappa}_n=\kappa(\widetilde{\mathbf z}_n), we construct Dt={(xn,yn):κ^n≤bt}\mathbf{\mathcal D_t=\{(x_n,y_n):\widehat{\kappa}_n\le b_t\}}, and restrict the phase-specific abduction space to St(yn)={z:z∧KB⊨yn,κ(z)≤bt}\mathbf{S_t(y_n)=\{\mathbf z:\mathbf z\wedge\mathrm{KB}\models y_n,\kappa(\mathbf z)\le b_t\}}. In the judicial sentencing task, ana_n is a coarse case-complexity descriptor supplied through domain-expert knowledge engineering, and ϕ(an,yn)\phi(a_n,y_n) returns its progression level. Cases are assigned to increasingly complex phases according to this level, with the final phase containing all cases. The effective phase dataset is Dt=DL∪{(xn,yn):ϕ(an,yn)≤t}\mathbf{\mathcal D_t=\mathcal D_L\cup\{(x_n,y_n): \phi(a_n,y_n)\le t\}}, where DL\mathcal D_L denotes the initially labeled subset. Correspondingly, the phase-specific knowledge sub-base admits candidate attribute vectors whose structural complexity does not exceed the current level. The online learner receives only the resulting phase identifier or eligibility mask, rather than the complete legal-attribute vector.

More generally, DPRP-ABL permits structural metadata to be supplied either directly by domain experts or indirectly by a frozen external expert, where access to \mathbf z_n^* is not a requirement of the learning framework. In strongly knowledge-dependent applications such as judicial sentencing, experts can provide coarse difficulty levels for previously adjudicated cases. In the digit addition benchmark, where digit identities are deliberately treated as unobserved concepts, we instead use frozen visual experts to avoid querying \mathbf z_n^*. We agree that the use of \mathbf z_n^* in the initial manuscript we submitted obscures the distinction between an idealized oracle constructor and our actual implementation, and we will revise the notation to define ana_n according to its actual source in each task.

Besides, our implementation does not set Dt=D\mathcal D_t=\mathcal D in early phases. Samples whose expert-assigned complexity exceeds btb_t are deferred and do not receive an early-phase pseudo-label. The final phase restores the complete dataset and knowledge base. These metadata determine when a sample becomes active and which knowledge sub-base is used, but they are never used as concept-level supervision or as additional numerical terms in the perception loss and candidate scoring. Subsequent phase advancement is determined automatically from perception-abduction agreement and the per-phase training budget, without consulting ground-truth concepts.

Moreover, all compared methods use exactly the same data splits. The distinction is that DPRP-ABL actively exploits the metadata to restrict the abduction search space and combines this restriction with distribution-prior-aware rectification for candidates. Thus, the improvement cannot be attributed to DPRP-ABL receiving additional annotations unavailable to the baselines.

We will expand Appendix H by adding details about dataset splitting, expert models, metadata construction, and progression design for both tasks.


(Hu et al., 2025b) Wen-Chao Hu, Qi-Jie Li, Lin-Han Jia, Cun-Jing Ge, Yu-Feng Li, Yuan Jiang, and Zhi-Hua Zhou. Curriculum abductive learning, 2025b.

AuthorsOpenReview ↗

Q2. Concerns about baselines.

Q2.1. Concerns about C-ABL and S1S_1 in our DPRP-ABL.

Thanks for your questions. S1S_1-only ablation is not identical to C-ABL. There are two main differences.

Firstly, that is phase construction. For example, in the implementation on the digit addition task, S1S_1 uses a carry-structured progression, i.e., the number of carry propagations, which conforms to the laws of human learning that have been claimed at P1 in Section 3.2. You can refer to Appendix H.1 to know more about the metadata construction and progression design. By contrast, C-ABL organizes curriculum by activating the digit-concept domain. They offline construct metadata by manually partitioning concepts into gradually expandable subsets until the final stage covers all concepts. Consequently, one of the differences is that every phase in S1S_1 can retain all digits, while the admissible reasoning structures and candidate space are progressively expanded. In the implementation of the judicial sentencing task, S1S_1 uses a legal case-structured progression. The implementation proceeds by progressively expanding according to the implicated attributes within the cases until all cases are included in the sixth stage. This is consistent with C-ABL. You can refer to Appendix H.2 to obtain more details.

However, what we cannot ignore is the second difference. That is the transition criteria among phases. The C-ABL transition criterion uses held-out concept-level accuracy in the digit addition task and F1 in the judicial sentencing. By contrast, in the implementation on two tasks, S1S_1 advances according to agreement between current perception predictions and revised pseudo-labels in abduction, as you can see in Eq. (6). Therefore, S1S_1 does not require ground-truth concept labels during online training.

This is the reason why we do not consider C-ABL as a comparison for the digit addition task in Table 1. Although concept labels in the held-out validation set are not involved in training as supervision signals contributing to gradient propagation, and they are used solely for evaluation, we think it might be unfair. We advocate that during the task processing, this information should remain entirely invisible. It does not affect the gradients in model training, nor should it influence the online training progression, and ground-truth concept labels should be accessible only offline.

To address this concern without introducing additional concept supervision during training processing, we have implemented a fixed-epoch transition adaptation of C-ABL using its concept-domain curriculum. This way is acknowledged by their authors. You can find details in their rebuttal on OpenReview. Results in the digit addition task are reported in Table 14 in Appendix I.1. Besides, for the judicial sentencing task, existing ABL methods capable of adapting to this task are scarce, so even if we may be disadvantaged due to unfair factors, we nonetheless proceed to compare in Table 4. However, a fair comparison can be judged by S1S_1-only ablation in Table 5, where the only difference between S1S_1 and C-ABL is the transition criteria. It can, to a certain extent, substitute for C-ABL in comparison with DPRP-ABL.

As you can see, results demonstrate that our progression design may be particularly beneficial as perceptual difficulty and abduction ambiguity increase, while DPRP-ABL additionally rectify majority dominance in candidate competition, alleviating skewness-induced bias in the perception-abduction loop.


Q2.2. Concerns about applying ImbNSL to the judicial sentencing task.

Thanks for your questions. Regarding ImbNSL, it assumes a known, fixed symbolic transition from discrete concept labels to discrete weak supervision and introduces marginal constraints at training or testing time (Tsamoura et al., 2025). However, the judicial sentencing task instead contains multiple coupled binary attributes, a continuous sentencing target, and a parameterized knowledge base whose coefficients are jointly learned. Therefore, ImbNSL is not compatible. We will clarify this applicability boundary in the revision.


(Tsamoura et al., 2025) Efthymia Tsamoura, Kai-Fu Wang, and Dan Roth. Imbalances in neurosymbolic learning: Characterization and mitigating strategies. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N.Chen (eds.), Advances in Neural Information Processing Systems, volume 38, pp. 152087-152123. Curran Associates, Inc., 2025.

AuthorsOpenReview ↗

Q2.1. Concerns about C-ABL and in our DPRP-ABL. Experiment results in Table 14.

Table 14. Comparison ''Avg±std'' with C-ABL on the digit addition task.

MethodsdMNIST z_ACC (↑)MNIST AUC (↑)MNIST PREM (↑)MNIST y_ACC (↑)SVHN z_ACC (↑)SVHN AUC (↑)SVHN PREM (↑)SVHN y_ACC (↑)CIFAR z_ACC (↑)CIFAR AUC (↑)CIFAR PREM (↑)CIFAR y_ACC (↑)
C-ABL10.9664±0.00560.9972±0.00050.9687±0.00480.9353±0.01060.8600±0.00740.9776±0.00160.8741±0.00490.7412±0.01280.3752±0.01020.7423±0.00750.5400±0.01810.1612±0.0069
DPRP-ABL (only S₁)10.9664±0.00930.9970±0.00160.9689±0.00830.9357±0.01740.9014±0.00590.9842±0.00090.9054±0.00460.8130±0.01020.3344±0.09770.7213±0.07130.4191±0.17320.1344±0.0648
DPRP-ABL10.9721±0.00160.9987±0.00030.9736±0.00120.9461±0.00270.9023±0.00020.9846±0.00050.9056±0.00080.8149±0.00050.3817±0.06190.7622±0.03910.5539±0.08040.1588±0.0450
C-ABL20.9789±0.00120.9985±0.00040.9797±0.00120.9199±0.00450.9024±0.00400.9874±0.00060.9069±0.00380.6643±0.01150.4825±0.01650.8171±0.00810.6093±0.01990.0616±0.0069
DPRP-ABL (only S₁)20.9780±0.00300.9987±0.00080.9790±0.00300.9174±0.01110.9176±0.00260.9875±0.00100.9187±0.00210.7099±0.00860.5943±0.08440.8706±0.03410.6898±0.04340.1403±0.0692
DPRP-ABL20.9789±0.00280.9990±0.00060.9799±0.00250.9206±0.01010.9262±0.00460.9889±0.00100.9270±0.00450.7360±0.01400.6176±0.03160.8865±0.01610.7068±0.01730.1515±0.0292
C-ABL30.9812±0.00170.9990±0.00020.9818±0.00170.8945±0.00940.9200±0.00480.9905±0.00080.9213±0.00460.6084±0.01870.5382±0.07580.8512±0.04160.6263±0.06150.0343±0.0196
DPRP-ABL (only S₁)30.9809±0.00330.9986±0.00060.9815±0.00330.8927±0.01770.9244±0.00090.9859±0.00080.9250±0.00090.6247±0.00280.6706±0.08320.9033±0.03090.7155±0.05500.1099±0.0483
DPRP-ABL30.9809±0.00150.9994±0.00020.9819±0.00120.8934±0.00780.9297±0.00180.9890±0.00090.9299±0.00190.6468±0.00750.7238±0.00950.9251±0.00840.7469±0.00200.1497±0.0117
Reviewer_PfwiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the problem of concept skewness in abductive learning (ABL), where majority-dominated predictions can introduce biased pseudo-label supervision and further amplify the underrepresentation of minority concepts. To address this issue, the authors propose Distribution-Prior-Rectified Progressive Abductive Learning (DPRP-ABL), which combines two complementary components: progressive expansion of knowledge sub-bases to reduce early candidate competition, and distribution-prior-aware rectification to mitigate majority-favoring biases in candidate selection. The paper also provides theoretical analyses of candidate-set contraction, score rectification, and ground-truth feasibility under certain conditions. Experiments on digit addition and judicial sentencing tasks demonstrate improvements in reasoning performance and concept recognition, supported by ablation studies evaluating the contributions of both components.

Critical strengths
  1. Reasonable motivation and a coherent methodological design.

The paper identifies a meaningful challenge in abductive learning: concept imbalance may be amplified through the interaction between perception predictions, abductive candidate selection, and pseudo-label feedback. The motivation is clearly explained, and the proposed two-component framework provides an intuitive approach to addressing this issue. Progressive knowledge scheduling reduces early candidate competition, while distribution-prior-aware rectification adjusts the preference toward majority concepts. These two components are logically complementary and form a coherent overall framework.

  1. Solid empirical evaluation with informative ablation studies.

The experimental evaluation covers both digit addition and judicial sentencing, including different perception backbones and knowledge-base settings. The method achieves competitive or improved performance across multiple settings, with particularly encouraging results on SVHN and judicial sentencing. The ablation studies further examine the individual contributions of the two components and generally support their complementary effects. Overall, the experiments provide meaningful evidence for the practical effectiveness of the proposed approach.

Critical weaknesses
  1. The methodological novelty is limited.

While the overall framework is well motivated, its two main components are closely related to existing techniques. Progressive knowledge-base scheduling has been explored in prior curriculum-based abductive learning methods, particularly C-ABL, while prior-based score adjustment is an established strategy in imbalanced learning. The contribution therefore appears to lie primarily in combining these techniques within the ABL framework rather than introducing a fundamentally new learning mechanism. A clearer discussion of the differences from closely related methods would help establish the significance of the contribution.

  1. The construction of progressive training phases requires further clarification.

Section 3.2 introduces task-specific structural metadata for determining the training phases, but the exact construction of this metadata is not sufficiently specified. In particular, the complexity measures in Appendix E depend on properties of concept assignments, such as carry counts in digit addition and the number of active attributes in judicial sentencing. Since concept labels are assumed to be unavailable during training, it is important to clarify how phase assignments are determined without relying on ground-truth concept information. This ambiguity does not necessarily invalidate the approach, but resolving it is important for assessing the validity and reproducibility of the reported results.

Questions for authors

Q1. Construction of progressive training phases: Could the authors clarify how the structural metadata and phase assignment function in Section 3.2 are instantiated for the two tasks? In particular, do the phase assignments require any information derived from ground-truth concept labels, such as carry counts or the number of active legal attributes? If not, how are these assignments determined using only information available during training?

Q2. Relationship to existing curriculum-based ABL methods.: Could the authors clarify the algorithmic differences between the progressive component of DPRP-ABL and C-ABL?

3 comments
AuthorsOpenReview ↗

Q1. Concerns about metadata construction and progression design.

Thanks for your questions. We clarify that the metadata are constructed entirely offline and are fixed before DPRP-ABL training begins. This general idea is inspired by the curriculum construction of C-ABL (Hu et al., 2025b). However, we adapt this idea to the structural characteristics of the two tasks according to our proposed principles P1-P3, which are detailed in Section 3.2. These principles organize the data and knowledge from simple to complex, making the progression compatible with the data-and-knowledge-driven learning paradigm of abductive learning and with an easy-to-hard learning process. The metadata is generated through an offline expert or domain-knowledge-engineering process, rather than by querying ground-truth concepts during online learning.

Specifically, for the digit-addition task, we use independently trained, frozen expert classifiers to recognize the input images and construct the carry-complexity metadata. The ground-truth digit concepts of the DPRP-ABL training samples are never queried by the progression constructor. Based on the expert predictions, equations are assigned to the no-carry phase, the at-most-one-carry phase, and the unrestricted final phase, which restores the complete training set and knowledge base. For the judicial sentencing task, the structural metadata are provided as part of the domain-expert knowledge-engineering process, where legal experts may organize previously adjudicated cases according to their structural complexity without exposing the complete attribute vector to the online learner. Cases are assigned to increasingly complex phases according to the expert-provided case-complexity level, with the final phase containing all cases.

In both tasks, the online learner receives only a phase identifier or eligibility mask, rather than the concept labels used to describe the sample. These metadata determine when a sample becomes active and which knowledge sub-base is used, but they are never used as concept-level supervision or as additional numerical terms in the perception loss, candidate scoring, or online phase transition decision.

We will expand Appendix H by adding details about dataset splitting, expert models, metadata construction, and progression design for both tasks.


(Hu et al., 2025b) Wen-Chao Hu, Qi-Jie Li, Lin-Han Jia, Cun-Jing Ge, Yu-Feng Li, Yuan Jiang, and Zhi-Hua Zhou. Curriculum abductive learning, 2025b.

AuthorsOpenReview ↗

Q2. Concerns about C-ABL and our DPRP-ABL.

Thanks for your suggestions. We agree on the importance of clarifying the relationship between our DPRP-ABL and C-ABL (Hu et al., 2025b). We acknowledge that the idea of the structure progressive component is inspired by the knowledge base curriculum introduced in C-ABL. However, we cannot ignore the differences between them in problem formulation, phase construction, transition criteria, and candidate scoring.

First of all, C-ABL introduces a knowledge base curriculum to reduce abduction space complexity and improve training stability. DPRP-ABL adopts this principle but studies a different failure mode, i.e., the majority-dominated feedback loop under concept skewness.

As for phase construction, for example, in the implementation of the digit addition task, our progression is organized by arithmetic structural complexity, i.e., by the number of carry propagations, which conforms to the laws of human learning that have been claimed at P1 in Section 3.2. You can refer to Appendix H.1 to know more about the metadata construction and progression design. By contrast, C-ABL organizes curriculum by activating the digit-concept domain. They offline construct metadata by manually partitioning concepts into gradually expandable subsets until the final stage covers all concepts. Consequently, one of the differences is that every phase in DPRP-ABL can retain all digits, while the admissible reasoning structures and candidate space are progressively expanded. In the implementation of the judicial sentencing task, our structural-complexity-based progression relies on the legal case complexity, based on which the metadata is constructed offline. The implementation proceeds by progressively expanding according to the implicated attributes within the cases until all cases are included in the sixth stage. This is consistent with C-ABL. Please refer to Appendix H.2 to obtain more details.

As for transition criteria among phases, for example, C-ABL transition criterion uses held-out concept-level accuracy in the digit addition task and F1 in the judicial sentencing. This subset of concepts is not participate in the backpropagation of gradients during training as a supervisory signal, it is used solely for evaluation. By contrast, in the implementation on two tasks, DPRP-ABL advances according to agreement between current perception predictions and revised pseudo-labels in abduction, as you can see in Eq. (6). Therefore, DPRP-ABL does not require ground-truth concept labels during online training.

Last but not least, in terms of candidate scoring, DPRP-ABL adds a component of distribution-prior-aware rectification, which estimates the concept marginals and calculates weight to rectify candidate utilities to counter majority-dominated pseudo-label selection. Although C-ABL contracts the abduction space but does not explicitly model or handle challenges caused by skewed concept distribution.

To support these distinctions empirically, in addition to what already existed in Table 4 and Table 5 in the initial manuscript we submitted, we have also added experiments for an explicit comparison in Table 14 in Appendix I.1 for the digit addition task. Results suggest that our progression design may be particularly beneficial as perceptual difficulty and abduction ambiguity increase, while DPRP-ABL additionally rectify majority dominance in candidate competition, alleviating skewness-induced bias in the perception-abduction loop.


(Hu et al., 2025b) Wen-Chao Hu, Qi-Jie Li, Lin-Han Jia, Cun-Jing Ge, Yu-Feng Li, Yuan Jiang, and Zhi-Hua Zhou. Curriculum abductive learning, 2025b.

AuthorsOpenReview ↗

W1. Concerns about the differences between our work and closely related methods.

Thanks for your important comment. We agree that curriculum scheduling and prior-based adjustment have precedents in curriculum-based abductive learning and imbalanced learning, respectively. However, DPRP-ABL is not merely a direct combination of them. Its contribution lies in identifying a failure mode in abductive learning and designing complementary strategies at its key stages.

Specifically, most traditional supervised imbalanced learning methods assume observed class labels. In abductive learning, concept labels are unobserved and are inferred through knowledge-consistent candidate selection before being used as pseudo-labels. Consequently, concept skewness may be amplified, where majority concepts are selected more frequently and are then further reinforced through subsequent training. We formulate this endogenous amplification as the skewness-induced biased feedback loop.

Our two components address different parts of this loop. Progressive knowledge scheduling controls which candidates compete during early learning, whereas distribution-prior-aware rectification controls how feasible candidates are ranked before pseudo-label selection. Unlike conventional supervised reweighting or post-hoc adjustment, our distribution prior is estimated without requiring complete ground-truth concept labels and is incorporated directly into abductive candidate scoring. This also distinguishes DPRP-ABL from C-ABL, whose curriculum was not designed to explicitly address concept skewness or majority preference in candidate scoring.

We think the significance of DPRP-ABL is that it reveals an important obstacle to applying abductive learning methods to naturally skewed data and develops a solution that intervenes within the feedback loop. This is important for extending abductive learning or more data-and-knowledge-driven learning methods to real-world applications where rare concepts remain practically critical. We will further expand Section 1 and the discussion of related work accordingly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We sincerely thank all reviewers for the careful reading and constructive comments. We appreciate these positive assessments, including:

  • A critical but rarely studied problem: We consider concept skewness and its amplification through the perception-abduction feedback loop to be a meaningful problem that is specific to abductive learning.

  • A coherent and targeted method: The dual-axis design is conceptually coherent, where progressive knowledge scheduling controls early candidate competition, whereas distribution-prior-aware rectification adjusts the relative preference among candidates.

  • A powerful empirical evaluation: There is the breadth and empirical strength of the evaluation, including the complementary ablations, the instantiations on both discrete and parametric knowledge bases, and the particularly encouraging results on two tasks.

We have provided detailed point-by-point responses to every concern and completed a series of additional analyses and experiments. We believe that we have successfully addressed most of their concerns. In our revised manuscript, we have incorporated the corresponding clarifications, methodological details, and new experimental evidence. We briefly summarize as follows:

  • Relationship to C-ABL and contribution positioning (reviewers Pfwi, rMMv, D8Xn, HMkh, and isAr): We clarify that DPRP-ABL differs from C-ABL in problem formulation, phase construction, transition criterion, and candidate scoring (the second comment for Pfwi). We have also added explicit comparisons among C-ABL, its inference-only post-hoc-adjustment variant, DPRP-ABL with only S1S_1, and the complete DPRP-ABL (Table 14 and Table 15 in Appendix I.1).

  • Direct evidence for minority learning and the feedback-loop mechanism (reviewer HMkh): We have added phase-wise candidate-set composition (Table 13), epoch-wise overall pseudo-label accuracy, minority pseudo-label precision and recall, and comparative training dynamics (Figs. 12-14 in Appendix I.1). We also separately report recall, precision, F1, and AUC for majority and minority concepts (Fig. 3 in Section 5.2 and Figs. 7-10 in Appendix I.1).

  • Robustness, prior estimation, and reproducibility (reviewers Pfwi, rMMv, D8Xn, HMkh, and isAr): We have added experiments under multiple skewness levels (Table 12 in Appendix I.1), evaluations about prior estimation values, estimation errors, and its update dynamics for both tasks (Table 10 and Table 11 in Appendix I.1 for task 1, Tables 16 and Table 17 in Appendix I.2 for task 2), and details of data splits, hyperparameter tuning, metadata construction, and phase progression (Fig. 15 in Appendix I.1 and analysis in Appendix H).

We believe these additions and clarifications comprehensively address reviewers' concerns and enhance the quality of our manuscript. All revisions are highlighted in blue in the updated manuscript for reference.

We once again thank all the reviewers for their hard work. As the discussion period is approaching its end, we would be very grateful if you could let us know whether our responses have addressed your concerns. If anything requires further clarification, please point it out. We look forward to any further comments and discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.