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

Detailed Comments:

Summary:

The paper presents a neural-symbolic framework that aims to integrate connectionist methods (such as large language models) with symbolic reasoning techniques to address complex, knowledge-intensive question answering tasks. The authors propose a modular architecture composed of a knowledge-oriented data manager, knowledge manipulator, reasoning planner, and reasoning conductor, and they emphasize the importance of human-in-the-loop interaction for continuous improvement. The ambition to blend the fast, intuitive processing of System 1 with the deliberate, step-by-step reasoning of System 2 (as framed by Kahneman’s cognitive theory) is an interesting angle.

Strengths:

- The paper clearly delineates the components of the proposed architecture and discusses how neural and symbolic methods might complement each other.

- Framing the integration in the context of dual-process theory (System 1 vs. System 2) provides a compelling narrative for why combining these approaches is desirable.

- The authors provide an extensive review of existing techniques and potential directions for integrating heterogeneous knowledge sources.

Weaknesses:

- The manuscript does not discuss in depth or compare its approach with recent advancements in reasoning with large models such as OpenAI o1/o3 or DeepSeek R1. These systems leverage step-by-step reasoning and test-time compute, concepts that align with the System 2 cognitive framework. A discussion of these works, including performance benchmarks or methodological differences, would significantly strengthen the paper. For instance, it would be helpful to explain why and when the proposed framework should be preferred to the large reasoning models.

- The work reads primarily as a position paper. While the conceptual framework is interesting, the submission lacks an empirical evaluation or a proof-of-concept demonstration. Without experimental results or case studies to validate the proposed architecture, the contribution remains largely speculative. An evaluation section with quantitative and/or qualitative results is necessary to substantiate the claims.

- If the intention is to address knowledge-based question answering specifically, the paper should cite and discuss related systems. For instance, the KBQA neuro-symbolic system described in “Leveraging Abstract Meaning Representation for Knowledge Base Question Answering” (https://aclanthology.org/2021.findings-acl.339/) is absent from the literature review.

Conclusion:
While the paper lays out an interesting framework and provides a detailed survey of potential techniques, it ultimately falls short of offering a concrete, validated contribution.

Major revisions are needed to:

- Incorporate a discussion and, ideally, comparative evaluation against recent state-of-the-art systems in reasoning with large language models.
- Provide empirical results or a demonstrable proof-of-concept to move the work beyond a speculative discussion.
- Update the literature review to include important works such as the KBQA system based on Abstract Meaning Representation.

Reviewer has chosen not to be Anonymous

Overall Impression: Good

Content:
Technical Quality of the paper: Good
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: Satisfactory
Level of English: Satisfactory
Overall presentation: Good

Detailed Comments:

Summary of the paper:
This paper proposes a general neural-symbolic architecture for knowledge intensive complex question answering. By reviewing the development of the artificial intelligence technologies in the past decades, authors emphasize the importance of combining the neural methods and the symbolic method to strengthen both paradigms. The proposed architecture includes four construction blocks, Knowledge-oriented Data Manager, Knowledge Manipulator, Reasoning Planner, and Reasoning Conductor. These four blocks carry distinct responsibilities and by combining these four blocks with possible human involvement in the loop, a more controllable and explainable neural-symbolic framework could be achieved.

Pros:
1. The paper is well organized to introduce the general neural-symbolic architecture, by reviewing the development of the AI technologies, introducing the the architecture in overview and detailed blocks, introducing related research and technique at the end.
2. The neural-symbolic architecture is general enough that could be achieved in diverse way, utilizing and combining the most advanced neural and symbolic methods.

Cons:
1. The introduction of the architecture is in a relative high and abstractive level. It would be better to include more concrete examples during the introduction, like the example “Who is taller, China’s most famous basketball player or the basketball player Lebron James?”.
2. In Section 3, a lot related research and technique towards the architecture is introduced. Though there are rich options for each block in the architecture, they could not be combined freely. For example, if the knowledge base is chose to be the knowledge-oriented data manager, then it is not likely to choose the textual corpora based solution for the knowledge manipulation operation realization. Thus it would be better to introduce some complete and possible framework following the proposed neural-symbolic architecture.

Questions and discussions:
1. The Knowledge manipulation module takes care of knowledge updating, which meanings to refresh and refine outdated information in the data structure. It is not clear whether the knowledge updating will result in updates of the data from the knowledge oriented data manager. Since in line 50 Page 8, it is said the knowledge-oriented data manager takes the responsibility of providing up-to-date knowledge. Then I would suppose the knowledge are up-to-date from the knowledge-oriented data manager block, and updates are not necessary in the other blocks. I’m a bit confused in this point.
2. Distinguishing between the steps assigned to the reasoning planner and the reasoning conductor is not always straightforward. For instance, the reasoning planner could generate a detailed plan for answering a given question, including steps such as identifying the most famous basketball player in China—an aspect currently handled by the reasoning conductor in the existing version. How should we determine which steps belong to the reasoning planner versus the reasoning conductor when implementing the proposed framework in real applications? More specifically, how can we effectively differentiate between the reasoning process and the planning steps?

Typos：
* Line 34 Page 1: … that can perceive, learn, reason, learn, and interact with …. (There is an extra ‘learn’)

Reviewer has chosen to be Anonymous

Overall Impression: Weak

Content:
Technical Quality of the paper: Weak
Originality of the paper: No
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: Average

Detailed Comments:

The authors propose a general architecture for reasoning tasks that involve heterogeneous knowledge sources and symbolic reasoning.

However, the description of the framework remains at a highly abstract and conceptual level. I struggle to see the novelty of the proposal or the advancements it facilitates. There is already a large body of work that integrates symbolic reasoning, LLMs, and other neural models in various ways. The authors do not clearly articulate how their approach differs from or improves upon existing methods.

The introduction presents a broad discussion inspired by Kahneman’s Thinking, Fast and Slow metaphor to frame the relationship between neural and symbolic methods in AI. However, it lacks a clear account of the paper’s contributions, its novelty, the key challenges, and relevant related work.

The description of the architecture consists of generic tasks that appear to be highly problem-specific. At such a general level, the paper does not provide any concrete guidance on how these tasks could be accomplished.

The section on techniques towards the proposed architecture reads more like a brainstorming list of various methods that are loosely related to the overarching idea. There is no clear explanation of how these techniques directly contribute to achieving the stated goals.

Minor Issues and Typos:
P1L34: The word learn is repeated twice.
P2L4: The claim that LLMs are recognized as AGI prototypes is questionable; I do not believe there is consensus on this.
P4L34: typically defines → typically define
P6L14: The meaning of the sentence "...surfers out of distribution tasks" is unclear.