Reviewer has chosen not to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes, but limited
Adequacy of the bibliography: Yes, but see detailed comments

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Needs improvement
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

This research provides a systematic mapping between neurosymbolic architectures and stages of bias mitigation. The topic is timely and a valuable contribution, as the integration of symbolic reasoning into fairness-aware AI has not been sufficiently explored. With improved structure, clearer differentiation from existing surveys, and enhanced presentation of figures and references, the work is publishable. The introduction appears to lack sufficient references and could benefit from clearer phrasing in several parts. The Bias Mitigation and Neurosymbolic Architectures sections show considerable overlap with previous studies, and the differences should be made more explicit. The remaining sections are suggested for minor revisions. The details are as follows:

1. Missing references:
　a. Trustworthiness aspects
Are these aspects your own interpretation? Or should they be cited?
　b. “Researchers in the field of neurosymbolic AI have proposed numerous architectures that incorporate the understandable, reasonable nature of symbols and statistical models that can handle noise and uncertainty.”
It is expressed as if there are clearly past studies, such as “proposed numerous architectures”, yet no such references are provided.
　c. “Most bias mitigation approaches encode constraints directly into the machine learning procedure and thus implement specific fairness notions for a distinct set of use cases.”
Add the case as a reference.
　d. “The biggest issue in symbolic systems is the grounding problem, i.e. to find an adequate mapping between the continuous real world and the assumed discrete world of the model”.
This is a well-known issue that has been the subject of past research, so a citation should be included.
Harnad, Stevan. "The symbol grounding problem." Physica D: Nonlinear Phenomena 42.1-3 (1990): 335-346.
2. Ambiguity and suggested rephrasing
　a. “Fairness stands out among these concepts, as this is an aspect of trustworthiness that ADM actually promises to enhance compared to human decisions”.
Clarify how fairness differs from the other components of trustworthiness, and in what way ADM systems are assumed to improve fairness relative to human judgment.
　b. “While bias detection queries whether data or a prediction satisfies a fairness constraint, bias mitigation employs fairness constraints on the data, the prediction model or the output”.
Does this mean “While bias detection queries whether predictions satisfy a fairness constraint, bias mitigation employs fairness constraints on the data and the models”?
　c. “i.e., are tied to one single formal definition of fairness”
What is meant by “one single formal definition”?
3. Clearly state the distinctions from other review studies in the same field. emphasise that this paper maps types of neurosymbolic approaches to bias mitigation techniques. This distinction should be explicitly linked to your listed contributions.
4. Clearly state the ultimate outcome or impact of the listed items in the contribution subsection.

1. Missing reference on “Many wide-spread notions of fairness focus on binary classification tasks with one binary protected attribute”.
For example:
Pagano, Tiago P., et al. "Bias and unfairness in machine learning models: a systematic review on datasets, tools, fairness metrics, and identification and mitigation methods." Big data and cognitive computing 7.1 (2023): 15.
2. Explain “b” in the equation in Group Fairness

1. Although well researched, the presentation is overly enumerative and challenging to follow, with unclear transitions to the subsequent section. The process of bias mitigation has already been reviewed in:
Hort, Max, et al. "Bias mitigation for machine learning classifiers: A comprehensive survey." ACM Journal on Responsible Computing 1.2 (2024): 1-52.
Rather than merely introducing it, please clearly articulate how your perspective differs from or extends it.

1. While new references have been added, the content appears largely identical to the following research:
Kautz, Henry. "The third ai summer: Aaai robert s. engelmore memorial lecture." Ai magazine 43.1 (2022): 105-125
Emphasize the distinct perspective of this manuscript.

1. “Chiappa (2019) proposed a method to adjust the output of a predictor to satisfy counterfactual fairness.” However, this is not listed in the references from Table 1.
2. Adding labels to each object in Figure 2 would be preferable to. The first model from left is SCM, the second is the neural network, and the third is the generated data. This is not explained in Section 5.3 (Neuro:Symbolic→Neuro Bias Mitigation).
3. Figure 3 would also be preferable to add labels to each object.

Reviewer has chosen to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes, but limited
Adequacy of the bibliography: Yes, but see detailed comments

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

This paper presents a well-structured and timely survey that maps neuro symbolic architectures to different stages of bias mitigation and the overall framing is sensible, with the taxonomy being useful for understanding how symbolic reasoning might support fairness interventions at various points in the ML pipeline. The paper fits well within the scope of the special issue and has potential as a reference or entry point for researchers working in this domain. That being said, several aspects would benefit from clarification / tightening before publication.

- A number of strong or normative claims in the introduction would benefit from clearer grounding or qualification. These include such statements as suggesting automated decision-making systems “promise” improved fairness compared to human decision-making, or that “most” bias mitigation approaches encode constraints directly into learning procedures. These claims are central to the paper’s motivation but currently under-supported. Narrowing the scope or adding references would strengthen the framing and avoid over-generalisation.

- Relatedly, while the paper’s main contribution is conceptual rather than technical, the novelty of the proposed mapping could be made more explicit. Several sections overlap in spirit with existing surveys on bias mitigation and neurosymbolic AI, and it would help to more clearly articulate how this work differs from or extends those prior efforts, particularly in relation to previously cited neurosymbolic fairness approaches.

- The treatment of fairness notions is generally accurate but remains quite high-level. Given the focus of the special issue, clearer signposting of which fairness definitions are assumed or in scope (e.g. group vs individual, statistical vs causal) and brief acknowledgement of known trade-offs would strengthen the conceptual rigour without significantly expanding the paper.

- From a presentation perspective, the taxonomy sections are informative but clearer transitions or short comparative summaries would improve narrative flow (vs the current enumerated list). Finally, there are a small number of minor typographical and notation inconsistencies (e.g. variable naming in some fairness definitions on page 5). These appear to be local and non-conceptual, but should be corrected for clarity and consistency.

Overall, this is a solid and relevant survey contribution and with clearer grounding of the key claims, a more explicit articulation of the conceptual contribution, and some small improvements to the presentation and precision, the paper would be suitable for publication.

Reviewer has chosen to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
Originality of the paper: Yes, but limited
Adequacy of the bibliography: Yes, but see detailed comments

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

The following comments can be addressed:
1-The paper is almost entirely conceptual and does not include any empirical evaluation to demonstrate that the proposed architectures work in practice.
2- No experimental comparison is provided against existing fairness-aware neural or neurosymbolic methods, making it impossible to assess practical benefit.
3- The authors should include at least one implemented architecture with quantitative results on standard fairness benchmarks (for example, COMPAS or Adult). Similar to the following paper, you could provide neumerical results and compare it with 3 existing approaches you mentioned like the following one. They did it on Adult, COMPAS and Lawschool datasets.
Heilmann, X., Manganini, C., Cerrato, M., & Belle, V. (2025). A neurosymbolic approach to counterfactual fairness. In 19th International Conference on Neurosymbolic Learning and Reasoning.
https://openreview.net/pdf?id=YZSDHz3Ydb

4- If the paper is intended as a survey, it lacks the depth, systematic methodology, and coverage expected from a review article.
5- The proposed taxonomy in Table 1 is nice, but it should be validated empirically.
6- The paper does not discuss how sensitive attributes are obtained or inferred in practice, which is critical for fairness systems.
7- The paper would benefit from a comparison table summarizing existing fairness methods versus proposed neurosymbolic approaches in terms of guarantees, cost, interpretability, and accuracy.
8- To be publishable as a regular research paper, the authors should add empirical validation, structured benchmarking, and reproducible experimental protocols; otherwise, the work should be reframed and significantly expanded as a systematic survey.