acceptodds
Under review as a conference paper at ICLR 2027

BSOR: Balanced Severity-Aware Ordinal Regression for Fine-Grained Misinformation Detection

Abstract

The spread of misinformation on social media has become a growing societal concern, making its automatic detection an important yet challenging problem. In practice, misinformation deviates from the truth to varying degrees, and fact-checking organizations therefore grade claims on ordered scales such as the six-level PolitiFact rating. Existing multi-class detection methods, however, treat these ordered grades as independent flat labels, overlooking the ordinal structure among them and consequently suffering from severe prediction deviation. To address this limitation, we propose Balanced Severity-Focused Ordinal Regression (BSOR), a plug-and-play output head that formulates multi-class misinformation detection as ordinal regression. BSOR builds on a conditional decomposition of the ordinal task and further introduces cost optimization based on misjudgment distance awareness and hierarchy equilibrium based on gradient normalization to align the training objective with the severity structure of the task, together with a KL-divergence-constrained soft-label regularizer for stability. Across two public datasets, replacing the flat head with BSOR consistently reduces MAE for nine general-purpose encoders and six representative detection methods, with maximum reductions of 6.99% and 7.94%, respectively. Through error-distribution analysis, we find that the improvement mainly comes from suppressing costly long-range mispredictions. Representation-level analysis further shows that the gains stem from a more sufficient learning of the ordinal relations implicit in the data. These results demonstrate that BSOR offers plug-and-play generality and achieves stable performance gains in fine-grained misinformation detection.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Misinformation Detection, Ordinal Regression, Fact-Checking, Natural Language Processing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ypNTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper treats six-level misinformation detection on the PolitiFact scale as ordinal regression rather than flat classification. The proposed BSOR is a plug-in output head built on CORN, with three additions: a loss weighted by misprediction distance, label-based sample weights that balance the subtasks, and a KL soft-label regularizer. On RDE and LIAR, replacing the flat head with BSOR reduces MAE for most of nine encoders and six detection methods.

Critical strengths
  1. The motivation is clear: fact-checking labels are ordered, and flat classification ignores this order.
  2. The head is simple to plug in, and it is tested across a wide range of encoders and detection methods.
  3. The error-distribution analysis shows that the gains mainly come from reducing predictions that are far from the true label.
Critical weaknesses

W1. The relevance of the method to current misinformation detection is unclear.

BSOR changes only the output head and loss of a classifier, and none of its components uses properties specific to misinformation (§4); the same head could be applied to any task with ordered labels. All experiments use encoder-based models without LLMs, such as BERT, RoBERTa, DeBERTa, dEFEND, and GEAR (Tables 2 and 3), and no LLM-based detector or LLM baseline is included. Current misinformation detection increasingly relies on LLMs, which may already capture the ordering of truthfulness labels, so it is unclear whether BSOR remains useful in this setting. A comparison with LLM-based methods, or an application of BSOR to an LLM-based detector, would help establish its significance.

W2. BSOR is not compared with existing ordinal regression methods.

The main results (Tables 2 and 3) compare BSOR only against flat classification heads. CORAL, which the paper cites (§2.2), is not compared, and the KL soft-label regularizer (§4.5) closely resembles soft ordinal labels [1], which is neither cited nor compared. The ablation (Table 4) suggests that most of the gain comes from CORN itself, as the authors also note (§5.4). For BERT-base, for example, CORN reduces MAE from 0.7402 to 0.7031, while the full BSOR reaches 0.6978. Since this ablation covers only three encoders on RDE, the added value of BSOR beyond standard ordinal regression remains unclear.

W3. The claim of consistent improvement may need qualification.

The abstract states that BSOR “consistently reduces MAE for nine general-purpose encoders.” This holds on RDE, but on LIAR, BSOR increases MAE for BiLSTM+Att (+6.56%) and WordAvgMLP (+2.95%) (Table 2). The authors explain that the flat baselines in these two cases collapse to a constant prediction (§5.2), which is reasonable, but the abstract’s wording does not reflect this exception. In addition, some gains are within one standard deviation (e.g., DeBERTa+LSTM: 0.5021±0.0044 vs. 0.4991±0.0057), and no significance tests are reported.

W4. The evaluation lacks class-level metrics, which leaves two claims unsupported.

The evaluation uses only MAE and Off-1 Accuracy (§5.1). Without accuracy, macro-F1, or per-class results, two points remain unclear. First, it is unclear whether BSOR trades exact accuracy for smaller errors: Figure 2(a) shows that exact-match predictions decrease for five of nine encoders on RDE and increase for the other four. Second, the paper argues that higher-label classes dominate training (§4.4), but provides no per-class evidence that this bias harms performance or that the corresponding component corrects it. That component also brings only small gains in the ablation (e.g., 1.0074 vs. 1.0073 for TextCNN+GloVe, Table 4).

The paper contains several typos, e.g., “Supplementar” (Figure 1, line 166), and missing spaces in “detection(Zhou” (line 094), “r_qi .The” (line 216), “assumption.To” (line 227), “bias.To” (line 261), and “methods.Six” (line 378).

[1] Raúl Díaz and Amit Marathe. Soft Labels for Ordinal Regression. In Proceedings of CVPR, 2019.

Questions for authors

See weakness.

6 comments
AuthorsOpenReview ↗

Response to W1 (1/2): Properties specific to misinformation and applicability to ordered-label tasks

Thank you very much for taking the time to review our paper and for the suggestions. Our responses and the corresponding revisions are given below. Due to space limitations, we summarize the theoretical analysis and the experimental improvements, so that the changes and results can be understood without reading the revised manuscript. If you would like further details, we have marked the corresponding locations in the revision.

W1 (1/2): No properties specific to misinformation, and applicability to any ordered-label task

① No properties specific to misinformation. The scientific goal of this paper is to make an initial verification of whether the ordinal regression paradigm itself is effective for fine-grained multi-class misinformation detection. For this purpose, we aimed to stay as close as possible to the native mathematical principles of ordinal regression, and we deliberately did not introduce overly complex customized network modules. Designing a completely novel and complex module from scratch would make the source of the gains hard to isolate: one could not tell whether the performance improvement comes from the ordinal regression paradigm itself or from the added complex architectural components.

Meanwhile, although we did not capture any misinformation-specific properties, we did adapt CORN to the practical needs of multi-class misinformation detection. This part was not clearly stated in the original manuscript, so the third paragraph of the Introduction in the revision now clarifies the relevance of BSOR to misinformation detection. Specifically, multi-class detection rates claims on an ordered scale, and the harm of a misjudgment grows with the level distance. This implies two ideal constraints on the training objective: the cost constraint, that the penalty on a misjudgment should increase strictly with the rank distance; and the stability constraint, that ordinal labels encode only the order of severity levels and not the importance of samples, so the loss aggregation under conditional decomposition should not structurally favor samples with higher label indices. In Appendix C, we further prove theoretically that neither flat classification nor vanilla CORN can satisfy these two constraints.

② The same head could be applied to any task with ordered labels. The two constraints fit ordered rating tasks whose cost is symmetric in direction and monotonically increasing with distance, and whose labels encode order rather than importance. Once the labels imply a direction-asymmetric cost, threshold-type decision boundaries, or the priority of the samples themselves, at least one of the two constraints no longer holds. We also note that fields such as medicine and credit rating do involve these two constraints, so our method is indeed not limited to misinformation detection; its actual effect on other tasks may depend on further properties and remains to be examined. This does not change our original motivation: these modifications were made to address misinformation detection.

The comparison with LLM-based methods is given in the next comment.

AuthorsOpenReview ↗

Response to W1 (2/2): Comparison with LLM-based methods

W1 (2/2): Comparison with LLM-based methods

③ Comparison with LLM-based methods. We agree that this comparison is necessary and have added it from two angles.

First, LLMs in direct question answering versus the best detection methods. We tested the direct question-answering capability of LLMs on both datasets: the locally deployed Qwen2.5-7B-Instruct with greedy decoding, and the API flagship model GLM-5; the output is restricted to a single digit from 0 to 5 corresponding to the six class labels. The controls are the best detection method on each dataset: dEFEND+BSOR on RDE and Speaker context+BSOR on LIAR. The results are as follows; the error distribution is the percentage of test samples in each off-by-k bin:

DatasetModel / MethodMAEoff-by-0off-by-1off-by-2off-by-3off-by-4off-by-5
RDEdEFEND+BSOR0.329884.26%5.03%7.04%0.82%2.78%0.05%
RDEQwen2.5-7B 0-shot0.634158.08%25.70%12.01%3.26%0.80%0.14%
RDEGLM-5 0-shot0.361568.24%27.72%3.73%0.27%0.02%0.02%
LIARSpeaker context+BSOR0.908346.84%29.31%13.48%7.17%2.88%0.23%
LIARQwen2.5-7B 0-shot1.537818.16%38.97%22.76%13.25%4.75%2.10%
LIARGLM-5 0-shot1.177728.76%39.67%20.42%7.64%3.20%0.31%

Direct question answering does not reach the level of the best fine-tuned detection methods on either dataset: on RDE, GLM-5 achieves an MAE of 0.3615 and Qwen2.5-7B achieves 0.6341, both higher than the 0.3298 of dEFEND+BSOR; on LIAR, GLM-5 achieves 1.1777 and Qwen2.5-7B achieves 1.5378, both higher than the 0.9083 of Speaker context+BSOR. The error distributions of the LLMs also shift to the right as a whole, and long-range errors remain severe. This shows that for fine-grained misinformation detection, which requires domain supervision and ordinal structure modeling, fine-tuned detection methods still have a clear advantage. This part is not reported in the main text because it is difficult to fit into the existing tables and due to space limitations.

Second, BSOR is equally plug-and-play for LLM backbones. We fully fine-tune Qwen2.5-1.5B and replace the Flat head with BSOR: on RDE the MAE decreases from 0.6236 to 0.6037, a relative reduction of 3.19%, and on LIAR from 1.3455 to 1.2445, a relative reduction of 7.51%; Off-1 Accuracy improves by 1.28 and 1.92 percentage points respectively, and both improvements are statistically significant. Its misjudgment-distance behavior is consistent with the pre-trained encoders. The Qwen2.5-1.5B results have been added to the encoder experiments in the revision; the direct question-answering comparison is new in this response. The details are given in Table 2 and Section 5.2 (page 7) of the revision.

In summary, the best existing detection methods outperform LLMs in direct question answering, and the BSOR head itself can be directly transplanted to LLMs with consistent improvements.

AuthorsOpenReview ↗

Response to W2 (1/2): Comparison with ordinal regression methods and the KL soft-label regularizer

W2 (1/2)

① Comparison with other ordinal regression methods. To demonstrate the advantage of our method over existing methods for multi-class misinformation detection, and the necessity of our adaptation-oriented improvements, we added comparison experiments with representative ordinal regression baselines, covering three families: CORAL (latent-variable thresholding), CDW-CE (cost-sensitive weighting), and EMD$^2$ (distribution matching). The experiments cover 6 settings: three encoders (RoBERTa-base, BERT-base, and DistilBERT) on the RDE and LIAR datasets. All methods are trained and evaluated with exactly the same training protocol as the Flat head. MAE is the mean of three random seeds; bold marks the best in each row.

DatasetEncoderFlatCORALCDW-CEEMD$^2$BSOR
RDERoBERTa-base0.59100.57830.60140.55980.5549
RDEBERT-base0.74020.72990.79170.69820.6978
RDEDistilBERT0.75830.76090.79150.74210.7286
LIARRoBERTa-base1.38561.17951.23151.29701.2897
LIARBERT-base1.31541.19431.20161.27771.2603
LIARDistilBERT1.32761.19431.24841.30111.2933

In the table, BSOR achieves lower MAE than Flat in all 6 settings, with an average reduction of 4.91% and an average Off-1 Accuracy improvement of 2.32%, verifying the effectiveness of BSOR over the flat classification baseline. Some ordinal baselines achieve lower MAE than BSOR on LIAR, which looks better; however, the new class-wise performance analysis shows that the gains of CORAL and CDW-CE come from biased prediction distributions: they concentrate predictions on the middle levels at the cost of abandoning the recognition of the boundary levels. The table below shows the per-class Recall results from that analysis: for RDE we excerpt the most biased setting (BERT-base), and for LIAR we give all three encoder settings. The Method column lists the four output heads Flat, CORAL, CDW-CE, and BSOR; values are percentages. The complete tables including EMD$^2$ and F1 are given in Appendix F:

DatasetEncoderMethodr0r_0r1r_1r2r_2r3r_3r4r_4r5r_5
RDEBERT-baseFlat58.534.738.830.174.055.5
RDEBERT-baseCORAL56.133.336.725.969.556.6
RDEBERT-baseCDW-CE24.538.932.736.863.830.3
RDEBERT-baseBSOR56.436.442.337.070.958.2
LIARRoBERTa-baseFlat28.323.330.723.735.118.5
LIARRoBERTa-baseCORAL1.86.342.547.730.31.3
LIARRoBERTa-baseCDW-CE0.08.036.147.326.80.0
LIARRoBERTa-baseBSOR21.726.129.329.633.720.2
LIARBERT-baseFlat12.031.922.427.745.418.0
LIARBERT-baseCORAL0.43.229.162.323.40.9
LIARBERT-baseCDW-CE0.07.130.154.431.70.0
LIARBERT-baseBSOR12.726.931.629.137.819.4
LIARDistilBERTFlat13.431.329.119.647.119.6
LIARDistilBERTCORAL1.87.141.148.329.51.9
LIARDistilBERTCDW-CE0.014.737.137.531.20.0
LIARDistilBERTBSOR12.727.126.034.132.516.0

The biased pattern is directly visible in the table. In all three LIAR settings, CORAL compresses the recall of the two boundary levels r0r_0 and r5r_5 to below 2 percentage points while piling predictions onto the middle levels; the recall of r0r_0 and r5r_5 for CDW-CE drops to zero in all three LIAR settings, and in the BERT-base setting on RDE it also compresses the recall of r0r_0 and r5r_5 by 34.0 and 25.2 percentage points relative to Flat. The apparent gains of both methods come at the cost of abandoning the recognition of boundary levels, which is not a desirable direction.

The details are given in Section 5.3 and Table 4 (pages 8–9) of the revision; the class-wise evidence is given in Appendix F (from page 33).

② The KL soft-label regularizer. We acknowledge that the KL soft-label regularization scheme used in this section closely resembles existing soft ordinal label methods, and the missing citation was a writing oversight. This module is directly reused from an existing scheme without innovative modification; its purpose is only to stabilize training and reduce mid-training performance jumps and convergence difficulty, and it brings no substantial performance gain. It is therefore only briefly mentioned in Section 4.5 and discussed together with the total loss, without being set as a separate ablation item. We have added the citation in the revision and a short note in the first paragraph of Section 4.5, clarifying that this regularizer comes from existing work and is not a contribution of this paper.

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

This paper studies six-level misinformation classification with ordered veracity labels. It proposes BSOR, a CORN-based conditional ordinal head augmented with distance-weighted likelihood, label-dependent loss balancing, and soft-label KL regularization. Experiments on RDE and LIAR replace flat heads in nine encoders and six adapted detection methods, reporting mostly lower MAE and fewer long-range errors, with ablations and representation analyses.

Critical strengths
  • Modeling the ordered veracity scale directly is a well-motivated solution strategy.
  • The authors present BSOR, a practical methodological benefit because it allows the same training design to be evaluated across different detection architectures.
  • Experimental results show that the output-stage modification is useful across many model pairs, including all six adapted detection methods.
Critical weaknesses
  • Threshold indexing starts at j=1, omitting the cumulative probability above the lowest rank and ending with an impossible above-maximum event. These equations prevent reproduction, although shifted code indices could mean a notation error rather than faulty experiments.
  • For six classes, the highest label joins at most five tasks, so weighting each by 1/6 sums to 5/6 rather than one (Equation 5). Moreover, equal loss weights do not equalize encoder gradients, and true-label weighting changes the fitted probability target; the claimed balancing and probability interpretations require correction or validation.
  • Severity weighting uses the current argmax class's distance, but inference uses cumulative thresholds and a median-type decision (Equation 4). These predictions can differ, and the weight changes discontinuously when the mode switches; no justification or comparison establishes an advantage over differentiable expected-distance or cumulative-distribution objectives for absolute rank error.
Questions for authors
  • What are the exact zero-based thresholds, masks, cumulative products, and class-probability reconstruction used in code?
  • Is the intended weight the inverse of the actual number of active tasks, capped at K-1?
  • How is the argmax-dependent weight differentiated, and why is it based on the mode rather than the inference-time median?
4 comments
AuthorsOpenReview ↗

Response to W1 and Q1: Correction of the indexing notation

Thank you very much for taking the time to review our paper and for the suggestions. Our responses and the corresponding revisions are given below. Due to space limitations, we summarize the theoretical analysis and the experimental improvements, so that the changes and results can be understood without reading the full revision. If you would like further details, we have marked the corresponding locations in the revision.

W1 and Q1: Correction of the indexing notation

The manuscript did contain indexing errors. The original notation had two problems: it omitted the cumulative probability above the lowest rank, P(y>r0)P(y > r_0); and it ended with a meaningless event, P(y>rK−1∣y>rK−2)=P(y>rK−1)≡0P(y > r_{K-1} \mid y > r_{K-2}) = P(y > r_{K-1}) \equiv 0. The inference formula inherited the same index shift. We have checked the code implementation and confirm that this deviation was solely a manuscript typo: the code uses zero-based indexing throughout. We have revised the entire manuscript to the zero-based indexing convention. The class-probability reconstruction, Equation (3), was already in zero-based form and required no change; this is also why the COMDA loss is unaffected by this revision. All experimental results remain unchanged. The specific modifications are listed below.

  • Section 4.2, Equation (1): the conditional probability definition is revised to fj(xi)=σ(wj⊤hi+bj)=P(yi>rj∣yi>rj−1)f_j(x_i) = \sigma(\mathbf{w}_j^\top \mathbf{h}_i + b_j) = P(y_i > r_j \mid y_i > r_{j-1}) with j=0,…,K−2j = 0, \ldots, K-2 (previously: j=1,…,K−1j = 1, \ldots, K-1).
  • Section 4.2, the paragraph after Equation (1): we add the boundary convention r−1≡−∞r_{-1} \equiv -\infty, so that f0(xi)=P(yi>r0)f_0(x_i) = P(y_i > r_0) is the unconditional probability of exceeding the lowest rank.
  • Section 4.2, the chain product: P(yi>rk)=∏j=0kfj(xi)P(y_i > r_k) = \prod_{j=0}^{k} f_j(x_i) (previously: ∏j=1kfj(xi)\prod_{j=1}^{k} f_j(x_i)).
  • Section 4.2, the monotonicity statement: the cumulative probabilities are described as monotonically non-increasing (previously: "naturally satisfy monotonicity").
  • Section 4.2, Equation (2): the predicted rank index is qi=∑j=0K−21{P^(yi>rj)>0.5}q_i = \sum_{j=0}^{K-2} \mathbb{1}\{\hat{P}(y_i > r_j) > 0.5\} (previously: ∑j=1K−11{P^(yi>rj)>0.5}\sum_{j=1}^{K-1} \mathbb{1}\{\hat{P}(y_i > r_j) > 0.5\}).
  • Section 4.3: the set of conditional probabilities is written as {f0(xi),f1(xi),…,fK−2(xi)}\{f_0(x_i), f_1(x_i), \ldots, f_{K-2}(x_i)\} (previously: {f1(xi),f2(xi),…,fK−1(xi)}\{f_1(x_i), f_2(x_i), \ldots, f_{K-1}(x_i)\}).
  • Section 4.4, Equation (5): both summation ranges are changed to ∑j=0K−2\sum_{j=0}^{K-2} (previously: ∑j=1K−1\sum_{j=1}^{K-1}).
  • Appendix B, Equation (8): the index is corrected to j=0,…,K−2j = 0, \ldots, K-2, consistent with Section 4.2.
  • Appendix B, after Equation (8): we add the boundary convention r−1≡−∞r_{-1} \equiv -\infty, so that f0(xi)=P(yi>r0)f_0(x_i) = P(y_i > r_0) is unconditional.
  • Appendix B, the definition of SjS_j: Sj={(xi,yi)∣yi>rj−1}S_j = \{(x_i, y_i) \mid y_i > r_{j-1}\} with Sj⊆DS_j \subseteq D (previously: {Sj}∈D\{S_j\} \in D).
  • Appendix B: S0=DS_0 = D contains all training samples (previously: "S1S_1 contains all training samples").
  • Appendix B, LcornL_{corn}: both summation ranges are changed to ∑j=0K−2\sum_{j=0}^{K-2} (previously: ∑j=1K−1\sum_{j=1}^{K-1}).
  • Appendix B, the inference chain product: P(yi>rk)=∏j=0kfj(xi)P(y_i > r_k) = \prod_{j=0}^{k} f_j(x_i) (previously: ∏j=1kfj(xi)\prod_{j=1}^{k} f_j(x_i)).
  • Appendix B: the cumulative probability sequence is revised to {P(yi>r0),…,P(yi>rK−2)}\{P(y_i > r_0), \ldots, P(y_i > r_{K-2})\} (previously: {P(yi>r1),…,P(yi>rK−1)}\{P(y_i > r_1), \ldots, P(y_i > r_{K-1})\}).
  • Appendix B: this sequence is monotonically non-increasing (previously: "monotonically increasing property").
  • Appendix B, the inference formula: qi=∑j=0K−21{P^(yi>rj)>0.5}q_i = \sum_{j=0}^{K-2} \mathbb{1}\{\hat{P}(y_i > r_j) > 0.5\} (previously: ∑j=1K−11{P^(yi>rj)>0.5}\sum_{j=1}^{K-1} \mathbb{1}\{\hat{P}(y_i > r_j) > 0.5\}).
AuthorsOpenReview ↗

Response to W2 and Q2: Weight sum, gradient equalization, and the fitted probability target

W2 and Q2: Correction of the weight sum; equal loss weights do not equalize encoder gradients; true-label weighting changes the fitted probability target

① The weight sum is 5/6 rather than 1. We acknowledge that this was a typo made during manuscript writing: the derivation and transcription of the formulas were not careful enough. In the original manuscript, the per-task weight 1/(yi+1)1/(y_i+1) makes the total weight of the highest rank only 5/65/6, which contradicts our claim that "the weights sum exactly to one." We have corrected the weight expression in the manuscript so that this claim holds exactly. The specific changes are:

  • Section 4.4: revised to "a sample with true rank yiy_i participates in min⁡(yi+1,K−1)\min(y_i+1, K-1) conditional subtasks in total ... and its loss term is multiplied by the weight 1/min⁡(yi+1,K−1)1/\min(y_i+1, K-1)."
  • Section 4.4, Equation (5): the per-term weight in the balanced loss LbalL_{bal} is updated to 1/min⁡(yi+1,K−1)1/\min(y_i+1, K-1).

② Equal loss weights do not equalize encoder gradients. HEGN is designed to normalize the total loss weight of each sample, mitigating the amplification effect min⁡(yi+1,K−1)\min(y_i+1, K-1), which is bound to the label order and introduced by the conditional decomposition. We note that equal loss weights do not guarantee equal gradient magnitudes at the encoder output. The actual gradient magnitude is determined by both the model's current predictions and the sample itself: correctly classified samples produce gradients close to zero, while misclassified samples produce larger gradients. We do not claim that gradient magnitudes can be fully equalized. This difference is an inherent property of all likelihood-based training paradigms and is a separate issue from the structural bias of ordinal decomposition that we aim to mitigate. To avoid misunderstanding, we have added an explanation in the first paragraph of Section 4.4 of the revision.

③ True-label weighting changes the fitted probability target. We agree that label-dependent weights change the fitted probability target, and we clarify this trade-off below.

Without reweighting, fjf_j as a maximum likelihood estimate is unbiased: it converges to the raw conditional frequency P(yi>rj∣yi>rj−1)P(y_i > r_j \mid y_i > r_{j-1}) of the training set. Statistically, this estimate does not deviate from the distribution of the training data, but it inherits the class imbalance of the dataset as it is. Taking the RDE dataset used in this paper (based on the PolitiFact six-level rating) as an example, the trustworthy label (True) accounts for only 15.9%, while the misinformation classes make up the vast majority. The raw conditional frequency of f0f_0 is close to 84%, so the model can easily learn the prior that "most statements are false" instead of learning discriminative features that separate truthfulness levels. More importantly, under CORN's conditional ordinal decomposition, misinformation samples join more conditional subtasks and naturally receive larger gradient contributions; the gradient signal of the minority True class is easily drowned out, so the model tends to simply predict statements as false, and the goal of fine-grained detection becomes hard to achieve. Therefore, the essence of this "unbiased" estimate here is that it faithfully inherits the distributional imbalance, which poses a serious risk for downstream fine-grained recognition.

HEGN introduces the weight wi=1/min⁡(yi+1,K−1)w_i = 1/\min(y_i+1, K-1) precisely to mitigate this structural amplification effect. Taking the PolitiFact six-level setting (K=6K=6, five conditional subtasks) as an example: a True sample joins one subtask with weight 1, while a Pants on Fire sample joins five subtasks with weight 1/5 on each. This setting compresses the contribution of high-rank misinformation samples to the total loss and balances the influence of different levels on parameter updates. The cost of this strategy, as you observed, is that fjf_j no longer fits the raw empirical conditional frequency; it fits the importance-weighted conditional frequency instead. In other words, the fitted target shifts. We have added an explanation at the end of Section 4.4 of the revision.

AuthorsOpenReview ↗

Response to W3 and Q3: Mode vs. median prediction rules, differentiation and discontinuity of the argmax weight, and comparison with differentiable expected-distance objectives

W3 and Q3: Mode vs. median prediction rules, differentiation and discontinuity of the argmax weight, and comparison with differentiable expected-distance objectives

① The inconsistency between the mode and median decision rules. When introducing the conditional ordinal regression head in Section 4.2, we aimed to faithfully follow the CORN framework and to mark the boundary of our modifications, so we kept CORN's original threshold-count inference rule. Our method intentionally differs from vanilla CORN in the choice between the mode and the median, but the COMDA section did not clearly state this difference. This was a lack of rigor in writing, and we have corrected it.

The severity weight of COMDA, wi=1+γ⋅∣y^i−yi∣w_i = 1 + \gamma \cdot |\hat{y}_i - y_i|, measures the severity of a single misjudgment. Here, the misjudgment refers to the deviation between the model's decision output y^i\hat{y}_i and the true label yiy_i. The mode (arg⁡max⁡\arg\max) is the level with the highest probability in the class distribution; it is the decision the model gives when it is forced to output a single prediction. The median is only a robust central statistic of the probability distribution; it marks the position that splits the probability mass and is not a decision the model outputs. Therefore, the severity loss should measure errors against the prediction the model actually commits to (the mode), not against a statistic the model never outputs.

Furthermore, because COMDA relies on y^i\hat{y}_i to compute the weight, we adjusted CORN's readout rule so that the prediction used for weight computation during training and the prediction output at inference are defined consistently. This adjustment only changes how the predicted label is read out from the recovered class distribution. It preserves the core mathematics of CORN, including the conditional probability decomposition and the chain-rule derivation of cumulative probabilities, and leaves the forward implementation and the definition of the conditional discriminators unchanged. This is also the core reason why COMDA uses the arg⁡max⁡\arg\max prediction rather than CORN's threshold-count median rule.

In the revision, Section 4.2 keeps the original description of CORN's native threshold-count (median) inference rule to mark the boundary of our modifications, and Section 4.3 (COMDA) now states that BSOR adjusts this inference rule.

② How the argmax-dependent weight is differentiated, and the discontinuity of the weight. argmax is indeed non-differentiable, but it appears only in the per-sample coefficient wi=1+γdiw_i = 1 + \gamma d_i, which is not on the gradient path. The implementation uses a straight-through estimator: in the forward pass, the hard prediction y^i=arg⁡max⁡mpi(rm)\hat{y}_i = \arg\max_m p_i(r_m) gives the misjudgment distance di=∣y^i−yi∣d_i = |\hat{y}_i - y_i| and the coefficient wi=1+γdiw_i = 1 + \gamma d_i; in the backward pass, wiw_i is treated as a constant (stop-gradient), and gradients flow only through the fully differentiable NLL term, i.e., ∂di/∂θ≡0\partial d_i / \partial \theta \equiv 0. argmax therefore only rescales the magnitude of each sample's gradient and does not change its direction. As you pointed out, the weight jumps when the mode switches; this discontinuity exists only in the forward value and never enters backpropagation. This implementation also has two favorable properties. First, wi∈[1,1+γ(K−1)]w_i \in [1, 1+\gamma(K-1)] is never smaller than 1, so no sample's gradient is blocked; gradients are only amplified according to the severity of the current misjudgment. Second, wiw_i is recomputed at every forward step and automatically decays back to 1 as y^i\hat{y}_i approaches yiy_i, with no extra scheduling. The implementation only requires a stop-gradient on the coefficient, introducing no additional parameters or structural changes. γ\gamma is set to 1.0 in the experiments.

③ Comparison with differentiable expected-distance or cumulative-distribution objectives. COMDA is itself designed to optimize the absolute rank error: its weight wi=1+γ∣y^i−yi∣w_i = 1 + \gamma|\hat{y}_i - y_i| is a monotonically increasing function of the model's absolute rank error ∣y^i−yi∣|\hat{y}_i - y_i|. The difference from a directly differentiable expected-distance objective therefore lies not in whether the absolute error is included, but in how it is implemented. COMDA keeps the likelihood term of each subtask and modifies the gradients only through a bounded sample weight wi∈[1,1+γ(K−1)]w_i \in [1, 1+\gamma(K-1)]. This preserves the distribution-fitting property of the likelihood while maintaining the rank-order constraint; when γ=0\gamma = 0, the form smoothly degenerates to vanilla CORN. A directly differentiable expected-distance objective replaces the likelihood terms and loses both properties. The advantage of this modular design has been validated by the ablation experiments on CORN vs. CORN+COMDA, and the misjudgment distance analysis in Appendix E also shows that the long-range misjudgments that dominate MAE are suppressed.

Thank you for the corrections and clarifications. My concern about how argmax-based weighting compares with differentiable distance-based losses remains. I will keep my original rating.

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

This paper studies fine-grained misinformation detection and addresses the limitation of conventional classification methods that ignore the ordinal relationships among truthfulness labels. It proposes BSOR, a plug-and-play ordinal regression head built upon the CORN framework. The method introduces misjudgment-distance-aware cost optimization, gradient-normalization-based hierarchy balancing, and soft-label regularization to improve ordinal prediction. Experiments on two public datasets, RDE and LIAR, show consistent MAE improvements across nine general-purpose encoders and six existing misinformation detection models, with a maximum reduction of 7.94%. Further analyses suggest that the improvements mainly come from reducing large-distance prediction errors and better exploiting the ordinal structure of the labels.

Critical strengths

1: Well-motivated problem formulation with practically valuable paradigm shift This paper is the first to formalize multi-class misinformation detection as an ordinal regression task. It accurately identifies the core limitation of flat classification methods — they disregard the inherent ordinal structure of labels and tend to produce costly cross-level mispredictions — which aligns closely with the real-world demand that larger judgment deviations incur greater harm in fact-checking scenarios. This work provides a concise, deployable new modeling paradigm for fine-grained misinformation detection. 2: Highly generalizable method with comprehensive and interpretable experiments As a plug-and-play output-head module, BSOR can be adapted to various encoders and specialized detection models without modifying the backbone network, resulting in very low engineering deployment cost. Experiments cover 9 general-purpose encoders and 6 domain-specific detection methods with thorough evaluation dimensions. The authors further clearly explain the sources of performance gain through error distribution decomposition and representation-level interpretability analysis. The conclusions are solid and credible, and the overall research completeness meets the standards of a top-tier conference.

Critical weaknesses
  • The technical novelty is relatively limited. The proposed method is built upon the existing CORN ordinal regression framework, while its main contributions are heuristic modifications to the loss function and training strategy. The paper does not provide theoretical analysis of the proposed mechanisms or sufficiently compare them with existing cost-sensitive and class-balanced ordinal regression methods. Therefore, it is difficult to assess the actual methodological advantage of BSOR over existing ordinal regression approaches.

  • The experimental evaluation is limited to two English political fact-checking datasets. The method is not evaluated in other misinformation domains, such as health or scientific misinformation, or in cross-lingual and multimodal settings. In addition, the paper does not provide sufficient sensitivity analysis or robustness evaluation under different label distributions. This limits the evidence for the general applicability of the proposed method.

  • The paper mainly reports improvements in MAE, but it remains unclear whether the proposed loss consistently improves classification performance across different ordinal categories. More detailed analysis of class-wise performance would help determine whether reducing large-distance errors comes at the cost of performance on specific classes.

Questions for authors
  1. How does BSOR compare with existing ordinal regression methods such as CORAL and cost-sensitive ordinal regression?

  2. How sensitive are the results to key hyperparameters and changes in the label distribution?

  3. How is the distance-aware loss implemented in practice, given that the predicted labels are obtained through a non-differentiable argmax operation?

6 comments
AuthorsOpenReview ↗

Response to W1 and Q1 (1/2): Technical novelty and supplementary theoretical analysis

W1 and Q1 (1/2): Technical novelty and supplementary theoretical analysis

Thank you very much for taking the time to review our paper and for your supportive review. Our responses and the corresponding revisions are given below. Due to space limitations, we summarize the theoretical analysis and the experimental improvements, so that the changes and results can be understood without reading the revised manuscript. If you would like further details, we have marked the corresponding locations in the revision.

① Technical novelty and contribution. We agree that the technical novelty of this work at the level of the ordinal regression backbone architecture is relatively limited. However, this design is a deliberate choice. The primary scientific goal of this paper is to verify whether the ordinal regression paradigm itself is effective for fine-grained misinformation detection. We therefore designed the method as closely as possible to the native mathematical principles of ordinal regression. If we introduced complex customized network components, it would be difficult to isolate variables, i.e., to tell whether the performance gain comes from the ordinal regression paradigm itself or from the added architectural modules. We chose CORN as the base framework precisely because its conditional binary decomposition gives a rigorous mathematical formalization of the ordinal structure, each subtask has independent parameters, the head is plug-and-play, and it avoids the shared-projection constraint that latent-variable methods such as CORAL impose on all categories. This fits the scientific goal of this study.

We have added the design motivation and the reasons for choosing CORN to the third paragraph of the Introduction and to the end of Section 4.2.

② Supplementary theoretical analysis. We recognize that the method lacked supporting theoretical analysis, so we have added Appendix C, "Theoretical Analysis of Constraint Alignment of BSOR for Fine-Grained Misinformation Detection." In this appendix, we first derive two ideal constraints that a training objective should satisfy for fine-grained misinformation detection: the cost constraint, that the penalty on a misprediction should grow strictly with its rank distance; and the stability constraint, that ordinal labels encode only the order of severity levels and not the importance of samples, so the loss aggregation under conditional decomposition should not structurally favor samples with higher label indices. We then prove that both the flat multi-class objective and the unmodified vanilla CORN baseline violate the two constraints by construction. Finally, we show that HEGN and COMDA restore the two constraints by construction respectively, re-aligning the training objective of BSOR with the task objective.

Specifically, Section C.2 proves that both the flat objective and the unmodified vanilla CORN baseline violate the two constraints by construction: the losses of both do not vary with the misjudgment rank distance di=∣y^i−yi∣d_i = |\hat{y}_i - y_i|, violating the cost constraint; a sample ii participates in min⁡(yi+1, K−1)\min(y_i+1,\,K-1) conditional subtasks, so its aggregated weight grows linearly with the label index, and even when the class distribution is perfectly balanced, the weight ratio between the highest and the lowest rank classes still reaches K−1K-1, violating the stability constraint. The two components of BSOR restore the two constraints by construction: HEGN normalizes the aggregated weight of each sample across all subtasks it participates in to exactly 1, shrinking the worst-case inter-class weight ratio from K−1K-1 to 11; the severity weight of COMDA, wi=1+γdiw_i = 1 + \gamma d_i, increases strictly with the rank distance. Constraint alignment is a necessary but not sufficient condition; the generalization gains are verified by the experiments in the main text.

On this basis, Section C.3 characterizes the mathematical nature of HEGN: it is equivalent to changing the loss of each sample over its participating subtasks from a sum to an average, so that the class-level contribution is proportional only to the class size. Section C.4 shows that COMDA is a sample weighting in the cost-sensitive learning sense: when γ=0\gamma = 0, it degenerates exactly to the unweighted original form, and the value of γ\gamma monotonically controls the ratio between the penalties of long-range and short-range misjudgments, which provides the theoretical basis for the hyperparameter sensitivity analysis.

The complete formal definitions, lemma proofs, and notation conventions are given in Appendix C (pages 15–18) of the revision.

The comparison experiments with ordinal regression methods are given in the next comment.

AuthorsOpenReview ↗

Response to W1 and Q1 (2/2): Comparison with ordinal regression methods

W1 and Q1 (2/2): Comparison experiments with ordinal regression methods

③ Comparison experiments with ordinal regression methods. To demonstrate the advantage of our method over existing methods for multi-class misinformation detection, and the necessity of our adaptation-oriented improvements, we added comparison experiments with representative ordinal regression baselines, covering three families: CORAL (latent-variable thresholding), CDW-CE (cost-sensitive weighting), and EMD$^2$ (distribution matching). The experiments cover 6 settings: three encoders (RoBERTa-base, BERT-base, and DistilBERT) on the RDE and LIAR datasets. All methods are trained and evaluated with exactly the same training protocol as the Flat head. MAE is the mean of three random seeds; bold marks the best in each row.

DatasetEncoderFlatCORALCDW-CEEMD$^2$BSOR
RDERoBERTa-base0.59100.57830.60140.55980.5549
RDEBERT-base0.74020.72990.79170.69820.6978
RDEDistilBERT0.75830.76090.79150.74210.7286
LIARRoBERTa-base1.38561.17951.23151.29701.2897
LIARBERT-base1.31541.19431.20161.27771.2603
LIARDistilBERT1.32761.19431.24841.30111.2933

In the table, BSOR achieves lower MAE than Flat in all 6 settings, with an average reduction of 4.91% and an average Off-1 Accuracy improvement of 2.32%, verifying the effectiveness of BSOR over the flat classification baseline. Some ordinal baselines achieve lower MAE than BSOR on LIAR, which looks better; however, the new class-wise performance analysis shows that the gains of CORAL and CDW-CE come from biased prediction distributions: they concentrate predictions on the middle levels at the cost of abandoning the recognition of the boundary levels. The table below shows the per-class Recall results from that analysis: for RDE we excerpt the most biased setting (BERT-base), and for LIAR we give all three encoder settings. The Method column lists the four output heads Flat, CORAL, CDW-CE, and BSOR; values are percentages. The complete tables including EMD$^2$ and F1 are given in Appendix F:

DatasetEncoderMethodr0r_0r1r_1r2r_2r3r_3r4r_4r5r_5
RDEBERT-baseFlat58.534.738.830.174.055.5
RDEBERT-baseCORAL56.133.336.725.969.556.6
RDEBERT-baseCDW-CE24.538.932.736.863.830.3
RDEBERT-baseBSOR56.436.442.337.070.958.2
LIARRoBERTa-baseFlat28.323.330.723.735.118.5
LIARRoBERTa-baseCORAL1.86.342.547.730.31.3
LIARRoBERTa-baseCDW-CE0.08.036.147.326.80.0
LIARRoBERTa-baseBSOR21.726.129.329.633.720.2
LIARBERT-baseFlat12.031.922.427.745.418.0
LIARBERT-baseCORAL0.43.229.162.323.40.9
LIARBERT-baseCDW-CE0.07.130.154.431.70.0
LIARBERT-baseBSOR12.726.931.629.137.819.4
LIARDistilBERTFlat13.431.329.119.647.119.6
LIARDistilBERTCORAL1.87.141.148.329.51.9
LIARDistilBERTCDW-CE0.014.737.137.531.20.0
LIARDistilBERTBSOR12.727.126.034.132.516.0

The biased pattern is directly visible in the table. In all three LIAR settings, CORAL compresses the recall of the two boundary levels r0r_0 and r5r_5 to below 2 percentage points while piling predictions onto the middle levels; the recall of r0r_0 and r5r_5 for CDW-CE drops to zero in all three LIAR settings, and in the BERT-base setting on RDE it also compresses the recall of r0r_0 and r5r_5 by 34.0 and 25.2 percentage points relative to Flat. The apparent gains of both methods come at the cost of abandoning the recognition of boundary levels, which is not a desirable direction.

The details are given in Section 5.3 and Table 4 (pages 8–9) of the revision; the class-wise evidence is given in Appendix F (from page 33).

AuthorsOpenReview ↗

Response to W2 and Q2 (1/2): Domain generalization and hyperparameter sensitivity

W2 and Q2 (1/2): Domain, cross-lingual and multimodal generalization, and hyperparameter sensitivity analysis

① Domain, cross-lingual, and multimodal generalization. The LIAR and RDE datasets used in this paper are not confined to political-election topics. For LIAR, purely political-process topics account for only 16.6% of all topic labels; the remaining 83.4% covers health care, economy and taxation, education, crime and justice, energy and environment, immigration, and social welfare. For RDE, politics-related topics account for a relatively high share of about 60.9%, and the remaining content covers public health, economy, education, science and environment, and entertainment. We have added a description of the topic distribution in Section 5.1 of the revision, with detailed topic proportions in Appendix D.1.

We acknowledge that the current experiments are conducted only on English plain text, and cross-lingual and multimodal settings have not yet been evaluated. Cross-lingual and multimodal misinformation detection are themselves two mature and large independent subfields with abundant dedicated studies. Adding cross-lingual or multimodal modeling modules to this paper would introduce extra confounding architectural variables, disturb the core conclusions, and shift the focus away from the goal of this work, namely verifying the feasibility and effectiveness of the ordinal regression paradigm for fine-grained misinformation detection. We therefore leave systematic cross-lingual and multimodal validation as important future work. We have added a discussion of these limitations in Section 6 of the revision.

② Hyperparameter sensitivity analysis. To examine whether the performance gain depends heavily on specific hyperparameter combinations, we added sensitivity experiments on the key hyperparameters of the training objective. Four representative models are selected, two from the encoder experiments (RoBERTa-base and TextCNN+GloVe) and two from the detection-method experiments (dEFEND and CompareNet). The four scanned hyperparameters are the severity focusing factor γ\gamma, the severity loss weight λsev\lambda_{sev}, the soft-label KL divergence weight λsoft\lambda_{soft}, and the soft-label distribution sharpness α\alpha. Each takes four levels around its default value: 0.25×, 0.5×, 1×, and 2×. All other settings remain unchanged and only retraining is done; the default level reuses the main-experiment results. In total, 52 configurations are evaluated on the full RDE test set.

The table below excerpts two representative settings, one from each experiment family: RoBERTa-base with the mildest fluctuations and CompareNet with the largest. The numbers of TextCNN+GloVe and dEFEND, as well as the Off-1 Accuracy trends of all models, are consistent with them.

ModelHyperparameter0.25×0.5×1× (default)2×
RoBERTa-baseγ\gamma0.56670.55110.55210.5647
RoBERTa-baseλsev\lambda_{sev}0.54670.56190.55210.5552
RoBERTa-baseλsoft\lambda_{soft}0.56360.55370.55210.5456
RoBERTa-baseα\alpha0.56230.56130.55210.5494
CompareNetγ\gamma0.32200.31870.31970.3273
CompareNetλsev\lambda_{sev}0.31870.30250.31970.2969
CompareNetλsoft\lambda_{soft}0.33340.33630.31970.2948
CompareNetα\alpha0.33000.33500.31970.3199

Measuring the fluctuation by the MAE range within the four levels relative to the default value, the fluctuations of all four hyperparameters are within 3.3% on RoBERTa-base and TextCNN+GloVe, within 8.5% on dEFEND, and within 7.1% on CompareNet except λsoft\lambda_{soft} at 13.0%; all levels are far from cliff-like changes. Notably, some non-default levels are even better than the default: for example, λsoft\lambda_{soft} at 2× gives an MAE of 0.5456 on RoBERTa-base, and λsev\lambda_{sev} and λsoft\lambda_{soft} at 2× give 0.2969 and 0.2948 on CompareNet, all below the respective defaults. This shows that the default configuration reported in the main tables is a conservative choice, not a value picked on the test set.

These results show that the gains of BSOR do not depend on specific hyperparameter combinations or fine-grained tuning, and the default configuration works stably over a wide range of values. This also corroborates the theoretical properties of COMDA in Appendix C.4: γ\gamma monotonically controls the ratio between the penalties of long-range and short-range misjudgments, and when γ=0\gamma = 0 the objective smoothly degenerates to the unweighted original form, so perturbations of the hyperparameters do not cause abrupt performance changes.

The details are given in the newly added Appendix H (page 46) of the revision.

The robustness evaluation under different label distributions is given in the next comment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: summary of revisions

Dear Area Chair and Reviewers,

Thank you for the substantial time and effort devoted to reviewing our paper and for the professional, detailed, and constructive comments. We are glad that the reviewers recognize the motivation of modeling the ordered veracity scale as ordinal regression, the plug-and-play generality of BSOR, and the comprehensiveness of the experiments. We have revised the paper based on all of the comments; the main work is summarized below, and each reviewer's specific concerns are addressed in the individual responses.

1. New theoretical analysis (Appendix C, a new section). To address the lack of theoretical support, we start from the task goal of fine-grained misinformation detection and formally define two ideal constraints: the cost constraint and the stability constraint. We prove that both the flat multi-class objective and the unmodified vanilla CORN violate the two constraints by construction (Lemmas 1 and 2), and that HEGN and COMDA restore them by construction (Propositions 1 and 2). We also characterize HEGN as changing a sum into an average, the cost-sensitive weighting of COMDA, the exact degeneration at γ=0\gamma = 0, and the monotone control of γ\gamma over the long-/short-range penalty ratio.

2. New comparisons with ordinal regression methods (Section 5.3, Table 4). To address the concern that existing ordinal regression methods were not compared, we added CORAL, CDW-CE, and EMD$^2$ in 6 settings (2 datasets × 3 encoders): BSOR achieves lower MAE than the flat classification head in all 6 settings, with an average reduction of 4.91% and an average Off-1 Accuracy improvement of 2.32%. The class-wise analysis further shows that the apparent advantage of some baselines comes from biased prediction distributions that abandon the boundary levels, while BSOR keeps non-zero recall for all six levels.

3. Ablation extended to all 15 settings (Section 5.4, Table 5). We extended the ablation from 3 to 15 encoder chains (9 on RDE and 6 on LIAR): the full BSOR attains the lowest MAE in all 15 settings, and the two modules target distinct weaknesses of conditional ordinal training and are complementary.

4. New LLM experiments (Table 2, Section 5.2). We added fully fine-tuned Qwen2.5-1.5B to the encoder experiments (now 10 encoders). Replacing the flat head with BSOR reduces MAE by 3.19%/7.51% on RDE/LIAR, both statistically significant, verifying that BSOR is equally plug-and-play for LLM backbones. Our response further adds direct question-answering comparisons with Qwen2.5-7B-Instruct and GLM-5, showing that zero-shot LLMs do not reach the level of the best fine-tuned detection methods.

5. New significance tests and class-wise analysis (Table 2, Section 5.5, Appendices E/F). All valid groups in the encoder experiments satisfy p < 0.05. The class-wise recall/F1 analysis, together with the misjudgment distance distribution, shows that the MAE reduction of BSOR comes from compressing long-range misjudgments rather than trading accuracy, and that the recognition of boundary levels is not sacrificed.

6. New hyperparameter sensitivity and label-distribution robustness experiments (Appendix H and the response). A sensitivity analysis of 4 representative models × 4 hyperparameters × 4 levels (52 configurations) shows that the default configuration is not the result of fine-tuning. A 10:1 imbalanced-resampling experiment and a six-class equal-quantity resampling experiment further verify the robustness of BSOR to label-distribution shifts and provide direct class-level evidence for the higher-label dominance argument in Section 4.4.

7. Important corrections. (i) The indexing notation has been unified to zero-based indexing, with the boundary convention added; (ii) the HEGN weight expression has been corrected to 1/min⁡(yi+1, K−1)1/\min(y_i+1,\,K-1), so that the weights of each sample sum exactly to one; (iii) we clarified that the inference readout of BSOR deviates from CORN's threshold-count (median) rule and uses the argmax, as noted in Section 4.3; (iv) the abstract has been qualified: steady reductions hold for all ten encoders on RDE and seven on LIAR, with exceptions due to degenerate flat baselines; (v) we added the citation for the KL soft-label regularizer (Díaz & Marathe, 2019) and stated that it is not a contribution of this paper; (vi) we added the topic-distribution description (Section 5.1) and a discussion of cross-lingual and multimodal limitations (Section 6); (vii) we corrected all typos and missing spaces.

We believe these revisions significantly improve the rigor, completeness, and credibility of the paper. We thank the AC and the reviewers again for their valuable time and comments, and we are very happy to discuss further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.