By Anonymous User
Review Details
Reviewer has chosen to be Anonymous
Overall Impression: Weak
Content:
Technical Quality of the paper: Weak
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: Satisfactory
Level of English: Satisfactory
Overall presentation: Good
Detailed Comments:
This paper surveys neuro-symbolic AI (NSAI) architectures and explores their potential in the context of generative AI. It revisits and extends Kautz’s taxonomy, and attempts to map various generative AI methods to these paradigms. The paper also proposes evaluation criteria for NSAI approaches and includes a case study on 4D printing as an illustrative application domain.
The authors address a timely and important question: how to bring structure to the fragmented space of neuro-symbolic AI, and how to connect it to the rapid developments in generative AI. This framing is worthwhile, as surveys that attempt to systematize the field can orient researchers who otherwise encounter disparate terminology and frameworks.
There are however several weaknesses and points for improvement:
- While the paper in the introduction claims to define and extend NSAI architectures, I do not see this reflected in the paper. The analysis is qualitative and descriptive, rather than providing deeper theoretical or empirical insights.
- The introduction of “fibring” as a new category is not well justified. As presented, it seems to resemble Symbolic[Neuro] with multiple neural components instead of one. Adding new categories without strong conceptual differentiation risks diluting the power of the classification.
- The section attempting to classify generative AI methods into NSAI paradigms is the most novel idea in the paper, but it falls short in execution. Several parallels feel superficial or misleading. Here are a few:
-- The claim that “XAI fails in the nested paradigm” oversimplifies a vast subfield and cannot be reduced to one category.
-- GANs are placed in the cooperative paradigm, yet they involve two neural networks without a symbolic component; it is unclear how this qualifies as neuro-symbolic.
-- Knowledge distillation is mapped to the compiled paradigm, though it is simply neural-to-neural compression; by the same logic, one could have mapped it to cooperative.
-- Fine-tuning and pre-training are treated as compiled neuro-symbolic methods, but these are standard neural training regimes without symbolic constraints.
-- Data augmentation is also listed under the compiled paradigm, but the example provided (synthetic data with logical rules) is conceptually distinct from typical augmentation.
Overall, these mappings appear forced, and the section does not achieve the promised contribution of establishing solid connections between NSAI and generative AI.
- The evaluation criteria introduced (generalization, scalability, interpretability, etc.) are reasonable dimensions, but the way they are applied in the large comparison table is not transparent. It is unclear on what basis the authors judged one architecture stronger than another on each criterion. Without explicit methodology, the evaluation reads as subjective opinion rather than reproducible analysis.
This paper makes a commendable attempt to structure the field of neuro-symbolic AI and to link it with generative AI. However, the claimed contributions are not convincingly realized: the taxonomy offers little beyond existing classifications, the NSAI vs. GenAI parallels are often superficial or inaccurate, and the evaluation lacks methodological grounding. The case study also overlaps with other published work by the same authors.
I recommend that the authors substantially revise the paper. In particular, they should:
- sharpen the taxonomy and avoid unnecessary category inflation,
- present a more rigorous and defensible mapping between NSAI and generative AI methods,
- clarify how evaluation judgments were made
