Reviewer has chosen not to be Anonymous

Overall Impression: Average

Content:
Technical Quality of the paper: Average
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: Average

Detailed Comments:

There is an improvement in clarity of the paper.
However I think the title is more general that the paper aims to be. The proposed architecture is specifically targeting Question-Answering (QA) tasks, and as such it is ok to relate to LLM mainly, but this is not a general blueprint for complex reasoning ( e.g. scene understanding is not discussed). There should be an indication in the title that this paper relates more to tasks that relate to QA (so as to justify the unbalanced set of references which are mostly referring to LLM and only occasionally to transformers for computer vision). This would also clarify how the Human In the Loop component is more related to humans on the other side of the QA interaction. However, if the human interaction is used to input additional correcting knowledge into the neuro-symbolic model this needs to be included in the blueprint and an account of approaches for knowledge injection (or some references to it) would be in order.

There are also a few statements without supporting references but I think that if the scope is narrowed properly and focused as suggested on QA (including change of title), that would be beneficial for the paper and for the audience.

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: Limited
Organization of the paper: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

Main comments:
I thank the authors for addressing my comments.

I have a couple of important comments about the new version (more detailed comments below):

1. The "complex knowledge-intensive reasoning task" is not defined. This is a critical fallacy of this paper, because despite the large number of qualifiers, it remains impossible to understand what is the scope of tasks that the authors have in mind. This needs to be explicitly discussed and some example tasks should at least be presented and motivated.

The running example used in the paper is an extremely unnatural question - it would be good to replace it with a task that can be more intuitively shown to be useful. The example also seems to be problematic from an objectivity standpoint, as "most famous" is a subjective category.

2. There are other NeSy architectures that are also inspired by Kahneman's dual system, such as LeCunn's JEPA [1] and others [2,3]. There needs to be a discussion of related work and an explicit comparison of this new architecture to existing ones.

[1] LeCun, Y. (2022). A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review, 62(1).

[2] Strannegård, C., von Haugwitz, R., Wessberg, J., & Balkenius, C. (2013). A cognitive architecture based on dual process theory. In Artificial General Intelligence: 6th International Conference, AGI 2013, Beijing, China, July 31–August 3, 2013 Proceedings 6 (pp. 140-149). Springer Berlin Heidelberg.

[3] Oltramari, A. A Path Towards High-Level Reasoning Through Cognitive Neuro-Symbolic Systems. (I think this one is under review at the same journal currently)

Abstract:

- it would be good to be explicit of what the "complex reasoning task" is.

Section 1:

- "learn, reason, learn" -> repetition
- it seems strange to say that the symbolic approach believes something.
- the introduction really needs more citations. Statements like the contemporary successor of symbolic approaches being knowledge engineering are not obvious, and the definition/scope of knowledge engineering has been debated as well.
- "competing and mutually reinforcing" - I am wondering whether the authors mean something else rather than "reinforcing"?
- figure 1's right picture seems off, as it doesn't show the same information as the others and it doesn't show what the caption says (model parameters).

Section 2:

- We credit knowledge manipulating, ... as fundamental skills of humans -> this needs a citation or an argument

Section 3:

- The components in §3 seem relevant, but it is hard to understand why are these necessary and complete. Pointers to prior reviews or surveys that support these classifications would enhance the credibility of these discussions and make them seem less like brainstorming.

Section 4:

- "many ... that indicates" -> "... indicate"