We would like to thank you for your valuable feedback, and we appreciate your recognition of our novel idea, the effectiveness of the proposed trust discounting method and stage-wise training algorithm, and its adaptivity to the existing evidential framework. We have carefully considered each of the concerns raised and have provided detailed responses below.

Experimental Designs Or Analyses:

(1) Our work primarily focuses on prediction conflicts, for example, the case shown in Fig. 1, where different views' prediction is different from others. We agree with you that explicit evaluation on datasets engineered to induce view conflicts is critical. However, existing literature lacks an ideal method for conflict simulation. For example, the only existing work, ECML, attempted to address the conflict issue, which randomly replaces original features with instances from other classes (e.g., substituting a "cat" with a "dog") to simulate the conflict issue. However, this approach has limitations: i). unrealistic conflicts: random replacements (e.g., swapping "cat" with "airplane") may create semantically implausible disagreements, unlike naturally occurring multi-view conflicts, the one we show in Fig.1. ii). distribution mismatch: artificially generated conflicts ignore view-specific feature distributions, potentially biasing evaluation results.

Instead, we propose an alternative conflict simulation method (acknowledging its imperfections, we move detailed results to Appendix D.7, Fig. 6–7 due to the space limitation). Our approach injects Gaussian noise into a randomly selected half of the views. The intuition is that as noise increases, corrupted views become less informative, approaching random guesses, while uncorrupted views retain (ideally) correct predictions, creating a controlled conflict scenario. We train models with noise injected data, ensuring the conflicts arise during the learning phase. As shown in Fig. 6–7, our methods (ETF/TF) outperform baselines in handling such conflicts.

(2) We agree that direct conflict measurement is important, and we regret that space limitations prevent us from including these results in the main text. In Appendix D.5 and Fig. 4, we measure pairwise view (or inter-view) prediction conflicts and also provide a concrete instance-level example of such conflicts in Fig. 5. Our methods TF and ETF builds upon TMC and ETMC, and as shown in Fig. 4, both our methods consistently and significantly reduce conflict ratios across all view pairs. For example, the prediction conflict ratio between the GIST view and the pseudo view of EMTC is 0.6, and our method ETF reduces it to 0.45, a 15% absolute improvement.

For the instance level case illustrated in Fig. 5, the original ETMC prediction was Label 4 (while the ground truth was Label 3). After applying ETF, the model correctly shifts higher belief mass to the true label (Label 3). Notably, this improvement is not limited to the combined view, it also enhances consistency in view-specific predictions (e.g., pseudo-view and View 2), demonstrating stronger agreement among views.

Essential References Not Discussed:

We appreciate the reviewer for highlighting the newly accepted ICLR 2025 work, which we recognize as concurrent research to ours and worthy of discussion in our submission. However, after thorough investigation, we were only able to identify one relevant paper titled ``Trusted Multi-View Classification via Evolutionary Multi-View Fusion". While the authors provide a GitHub link, we note the implementation appears unavailable at this time.

Upon careful analysis, we find their work orthogonal to ours: their primary contribution focuses on enhancing pseudo-view quality through evolutionary fusion, whereas we propose a fundamentally different approach for handling prediction conflict via Trust Discounting. That said, we acknowledge potential synergies - our ETF method could potentially incorporate their evolutionary architecture by: i) integrating our referral ENN to further refine the pseudo-view generated by their evolutionary fusion, and ii) maintaining our core innovation in conflict resolution while benefiting from their improved pseudo-view feature. This complementary combination might yield additional improvements, though we leave such exploration for future work given the current unavailability of their implementation.

In the revised manuscript, we will explicitly cite and discuss the recommended paper, highlighting its relevance and clarifying how our work differs from or builds upon it.

We hope our responses effectively address your concerns and appreciate if you can reconsider and improve the evaluation.

We sincerely appreciate the reviewer's valuable feedback and their recognition of: our research motivation, the novelty of proposed trust-based discounting mechanism and effectiveness of training algorithm. We have carefully addressed each of the concerns raised with detailed responses below, and we hope these clarifications satisfactorily answer all questions.

Questions For Authors:

(1) We adopt the terminology from the Subjective Logic book, distinguishing between referral opinions and functional opinions generated by different types of Evidential Neural Networks (ENNs). Specifically, i) Referral opinions serve as reliability assessments, indicating whether a prediction should be trusted, ii) Functional opinions directly support decision-making based on their own evidence. We confirm your observation that these are generated by two distinct sets of evidence network parameters, and there are still some key differences worthy of highlighting.

• As noted in line 266 and line 270, the referral ENN and functional ENN process different input features, referral ENN takes one more input which is the belief masses of functional opinion;
• Their roles are also different, as in Eq. 5 (Trust Discounting Mechanism) – referral opinions modulate trust, while functional opinions provide evidence for decision making;
• Referral opinions are always derived from beta distributions (Fig. 2), quantifying the reliability of their associated functional opinions. However, functional opinions may follow either: 1) Dirichlet distributions for multiclass classification or 2) Beta distributions for binary classification.

(2) We appreciate this important question. Let us analyze potential limitations from both global and local perspectives.

• From a global perspective, the evidential multi-view classification framework relies on late fusion. This means that there is no early interaction between views during feature extraction. Although effective in many cases, this design might limit knowledge integration when different views provide complementary but partial information about the complete pattern.
• From a local perspective, our Trust Discounting (TD) module has a potential limitation: it may still use referral opinions with high uncertainty. Since high uncertainty suggests lower reliability, this indicates that our current trust adjustment mechanism could benefit from more fine-grained handling of such cases. We acknowledge this as an area for future improvement.

Weaknesses:

(1) The difference between ECML and our work can be summarized as follows,

• Different Conflict Resolving Mechanism: our method uses a trust discounting module to modulate the trust on the functional opinions, while ECML uses a loss function to harmonize different views' functional opinions (this is already mentioned in the related work section).
• Our method is built upon TMC and ETMC, so using the same belief fusion method, which is the Dempher-Shafer rule for combining different views from opinion perspective. However, ECML uses a different, evidence averaging based method to fused opinions.
• To keep the effectiveness of the introduced TD module, we proposed stage-wise training algorithm, which is also different from ECML.

(2) Thank you for raising this important point. A similar concern regarding Experimental Designs or Analyses has been addressed in detail in our response to Reviewer ngw9 (please refer to our first response statement to Reviewer ngw9).

(3) Please refer to our reply to your same question above

(4) Thank you for raising this important point. Our method consistently outperforms existing baselines across all six evaluated datasets. We acknowledge that in some cases, the improvements might appear marginal, which we attribute primarily to the following reasons:

• Limited improvement space: For datasets such as Handwritten, all compared methods already achieve accuracy levels above 99%, leaving minimal room for further substantial gains.
• Intrinsic difficulty of certain datasets: For challenging datasets like Caltech, existing state-of-the-art methods typically achieve around 94% accuracy. Despite this intrinsic difficulty, our method successfully pushes performance beyond this threshold into the 95% range, demonstrating effectiveness even under constrained conditions.
• Baseline instability and fluctuations: On datasets such as CUB, baseline methods exhibit performance instability, fluctuating notably between 90% and 93% accuracy. In contrast, our approach demonstrates clear superiority by consistently achieving around 94% accuracy, indicating greater robustness and stability.

Additionally, it's important to highlight that on datasets like Scene, our method achieves a substantial improvement of roughly 4-5%, while even in the least improved scenario (PIE and HMDB dataset), we still observe a meaningful increase of above 1%.

We sincerely appreciate the reviewer's recognition of our work's strengths:

1. the clear organization and presentation of our manuscript, 2) the comprehensive literature review demonstrating deep domain understanding 3) the methodological innovation in our trust-based fusion approach, and 4) the extensive experimental validation demonstrating effectiveness. We have carefully addressed each of the concerns raised and provide detailed responses below. We hope these clarifications fully resolve any remaining questions about our work.

Weaknesses:

1. Thank you for raising this important point. A similar concern regarding Experimental Designs or Analyses has been addressed in detail in our response to Reviewer ngw9 (please refer to our first response statement to Reviewer ngw9).

2. Regarding network architecture, we maintain consistency with established approaches (TMC, ETMC, ECML, etc.) for fair comparison:

   • For the six vector-based datasets, we use identical architectures: a single-layer network with Softplus activation to generate functional opinions.
   • For the Food101 dataset, we employ standard pre-trained models (ResNet50 for images, BERT-base-uncased for text) as feature encoders (as noted in Section 4.4), and referral and functional networks share these encoders on each view. However, each ENN maintains its own output layer (linear + Softplus) to ensure specialized processing while leveraging common feature representations.

3. Please refer to our response to Reviewer icFc, we will include the limitation our work in the revision.

Other Comments Or Suggestions:

We thank you for pointing this out. We believe that clarifying our innovation will make it clearer.

1. While the original subjective logic framework proposes a global trust-discounting mechanism, we extend it to operate in an instance-wise manner. This critical advancement enables adaptive handling of varying conflict patterns across different samples, significantly enhancing the framework's flexibility.
2. We introduce a novel stage-wise training approach that is essential for keeping the effectiveness of our Trust-Discounting (TD) module. Based on our study, simply incorporating the TD module into existing training frameworks (e.g., ETMC's approach) would actually degrade performance. Our designed training strategy ensures stable optimization and reliable conflict handling.

Questions:

We acknowledge that the learned weight p reflects the inherent patterns in the training data. Specifically, when certain views exhibit consistent conflicts, the framework will assign them lower weights. However, our design is fundamentally instance-wise, meaning it remains effective even when conflicts lack clear trends across the dataset. For example:

1. in our controlled experiments (Appendix D.7, Fig. 6–7), we artificially introduced random conflicts by corrupting 50% of views with noise—a scenario without view-specific bias.
2. under these conditions, such as our method (ETF) still outperformed baselines like ETMC, demonstrating its robustness to sporadic conflicts. This validates that the framework adapts not only to systematic conflicts but also to instance-specific conflicts. We appreciate the opportunity to clarify this point.

We thank you for your feedback and below is our response to your questions,

1. Regarding AUROC usage: We clarify that our use of AUROC differs fundamentally from TMC [Han et al., 2022]. While TMC employed AUROC to measure label prediction accuracy, we follow the established uncertainty evaluation framework of [Filos et al., 2019], where AUROC assesses uncertainty calibration - specifically, whether the model's confidence in its predictions correlates with their correctness. In other words, it measures the rate of referring the least confident predictions to human experts.

   This distinction is important because:

   • As noted in TMC's results (Table 1), AUROC for label prediction consistently shows high scores ($>$95%) even when accuracy is modest (67.74% on Scene15, 65.26% on HMDB), suggesting limited discriminative value
   • Subsequent works (ECML, CCML, TMNR) have consequently abandoned this usage
   • Our application evaluates a fundamentally different capability: the model's ability to identify unreliable predictions through uncertainty scoring.

2. Writing logic: We appreciate the suggestions regarding paper organization. While we will carefully consider these recommendations for future work, we believe the current structure most effectively presents our technical contributions and results.

3. Wording improvements: We thank the reviewer for their thoughtful suggestions on phrasing and will incorporate appropriate refinements while maintaining the paper's technical precision.

4. Regarding the question: We appreciate your close examination of our results. There appears to be a misunderstanding regarding the relationship between Tables 1 and 3, which we would like to clarify:

   • Table 1 demonstrates the limitation of common belief fusion methods (like BCF [Han, 2022]) where, as we stated in Line 190, "the fused opinion is expected to be higher than those of each individual views" - this represents the problem scenario we aim to solve
   • Table 3 presents results after applying our Trust Discounting mechanism, which addresses the conflict issue. The values shown (0.08 < 0.10 < 0.42 < 0.76) are: a) expected outcomes of our method, b) demonstrate proper handling of conflicts, and c) represent an improvement over the Table 1.