Reviewer has chosen to be Anonymous

Overall Impression: Average

Content:
Technical Quality of the paper: Average
Originality of the paper: Yes, but limited
Adequacy of the bibliography: No

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

Detailed Comments:

The author seems to be taking a "Band-Aid" approach towards developing cognitive models/systems: patching up some holes here and there, without fundamentally thinking (or re-thinking) the issues of cognition and models of cognition. LLMs and generative AI provide us with a golden opportunity to think or re-think seriously fundamental issues in this regard. I can point the author to one (of many other places) where such serious thinking takes place --- for example, a recent volume titled: Cambridge Handbook on Computational Cognitive Sciences, with many chapters dealing with many different issues.

Author's discussion in Section 4 about actual details of integrating neural and symbolic components is confusing. It is apparent that author has not yet given it enough thinking. This shows. The author should make this explicit to begin with.

Another point is the selection of ACT-R by the author is curious. What about other well-known cognitive architectures (especially those that are meant to be psychologically realistic models of the human mind), such as EPIC, CLARION, NANGO/SPA, and so on? Some of them are more psychologically realistic, or are neuro-symbolic to begin with, or both? To some of them, being neuro-symbolic is a given (e.g., CLARION) and it can lead to many interesting capabilities. The author should, at a minimum, discuss these possibilities and their potential advantages.

Reviewer has chosen to be Anonymous

Overall Impression: Average

Content:
Technical Quality of the paper: Weak
Originality of the paper: Yes
Adequacy of the bibliography: Yes

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

Detailed Comments:

A Path Towards High-Level Reasoning Through Cognitive Neuro-Symbolic Systems
Alessandro Oltramari

This paper propose to perform high-level reasoning (including commonsense reasoning )by augmenting cognitive architectures with external neuro-symbols components.
The author argued in this paper that compressing KG into low-dimensional spaces has limitations in high-level reasoning and common sense reasoning which both are required for complex tasks.
In this position paper they propose a conceptual framework for cognitive neurosymbolic reasoning systems centred on a cognitive architecture (based on standard model of mind). In this paper they show how this can be done with the well-known ACT-R cognitive architecture.

This is a very interesting approach, and interesting topic for the journal.
However the proposed cognitive architecture is rather superficial discussed. It deserves much more space in the paper. My suggestion is to shorten section 3 and to extend section 4.

- Figure 1 is not needed.
- The motivation part in section 3 is too long. I think it is much more interesting to go more in depth in section 4 the proposed architecture (=main contribution of the paper).
- figure 2: mention in the caption that the blue boxes are the ACT-R modules, and the green boxes the neurosymbolic extensions.
- suggestion: use numbers for the links that are described on page 5, such they can be easily connect to the figure 2 by including those numbers in the figure.
- knowledge —> neural (page 5). Clarification question: This is channelled via the declarative memory to the ACT-R buffer?
- Is the procedural model the “pattern matching and production execution” components in fig. 2.
- Production system = procedure model?
- page 5, r33. “Such feature” to which feature is referred, and why does this make the system efficient. (efficient with respect to…)

- language: doesn’t —> does not, don't --> do not

Reviewer has chosen 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: Limited
Organization of the paper: Needs improvement
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

The paper addresses the topic of the journal and reads well overall as a position paper for the inaugural invited issue. The author should consider to address the following issues:

- At the conceptual level, it seems that the underlying mechanism of ACT-R is already neuro-symbolic to some extent. Therefore, integrating ACT-R with neuro-symbolic reasoning systems does not sound very fitting. Perhaps there is a better way to phrase this.

- Injecting knowledge into a neural network (or simply “knowledge-infusion") can be done in many ways, e.g., by encoding the knowledge as a loss function (also known as "semantic loss”). A proper discussion of different methods for knowledge-infusion in the related work section would be helpful.

- Many references are outdated. For knowledge embeddings please refer to recent references (see below). There are also several old arXiv references that could be potentially published somewhere else by now.

- The Motivation section (Section 3) is not well supported by references. All references for criticizing the limitations of LLMs are chosen from pre-ChatGPT era (i.e., before 2023), which were valid at that time but may not be valid any more now. Recent references that pinpoint the limitations of ChatGPT and GPT-4 would be necessary to make the Motivation section stronger.

- The procedural module and inference engine are not visible in Figure 1.

- How does the proposed cognitive neuro-symbolic system overcome the limitations in other cognitive architectures, e.g., LeCun’s cognitive architecture [47]?

Knowledge Graph references:
- S. Ji, S. Pan, E. Cambria, P. Marttinen, and P. S. Yu, “A Survey on Knowledge Graphs: Representation, Acquisition, and Applications,” IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 2, pp. 494–514, Feb. 2022, doi: 10.1109/TNNLS.2021.3070843.

- A. Hogan et al., “Knowledge Graphs,” ACM Comput. Surv., vol. 54, no. 4, p. 71:1-71:37, Jul. 2021, doi: 10.1145/3447772.

Reviewer has chosen to be Anonymous

Overall Impression: Bad

Content:
Technical Quality of the paper: Bad
Originality of the paper: Yes, but limited
Adequacy of the bibliography: Yes

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

Detailed Comments:

The paper proposes a methodology to use cognitive architectures, augmented by neuro-symbolic components, to achieve high-level reasoning in AI. The topic is very timely and important. However, the paper is very weak in its proposal.

First, it does not give any description of the ACT-R architecture, which is at the base of the proposed method. Even if this is background work, the paper should be self-sufficient and describe the modules of this architecture, their interactions, and the overall functioning.

Second, the additional neuro-symbolic models are vaguely described at a very high level. This is not enough to assess the value of the proposal.

Third, some examples of the functioning of the proposed architecture should be provided. These examples should show how some of the missing reasoning capabilities are instead provided by the proposed architecture.