acceptodds
Under review as a conference paper at ICLR 2027

Toward a Unified Mathematics of Concepts

Abstract

Concepts are commonly defined as abstract, compact representations of knowledge and treated as basic units of intelligent behavior. Yet, cognition, psychology, and AI lack a shared mathematical language for them. Modern systems represent concepts as vectors, distributions, symbols, graphs, and other structures, but these formalisms are typically treated as competing rather than as solutions to a common problem. We propose an operation-based view that evaluates mathematical frameworks by the conceptual operations they support, identifying thirteen operations (including similarity, composition, generalization, and grounding) that recur across cognition, psychology, and AI. We show that ten frameworks embody distinct commitments to concepts as self-contained content, relational structure, or evolving process, and that these commitments determine which operations each supports naturally. For example, vector-based models facilitate graded similarity and generalization but struggle with explicit composition, whereas symbolic models support composition but offer but generalize poorly. No single framework we examined naturally supports all operations without extension. We test this account empirically using categorization as a case study, operationalizing nine theories on the same items against human judgments. Despite addressing the same conceptual question, the theories produce different procedures and results, demonstrating that mathematical commitment shapes what a theory can explain. We call for hybrid formalisms that treat content, relation, and process as jointly primary.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Concepts, Mathematics, Human Modeling, LLM Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ZWpaOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper claims to investigate how different mathematical frameworks operate on concepts in supporting various reasoning patterns. It presents the work as lists of reasoning patterns, mathematical frameworks, concept representation methods. Then it centers on concept classification as a short case study with examples, ignoring all the complexities described in the seminal works it is citing. The authors do not provide sufficient analysis to support their claims of systematically describing a broad set of methods, nor do they propose a new system of their own that would justify the title.

Critical strengths

The authors collect seminal works in their references, this would be a good starting point for a literature review of reasoning methods.

Critical weaknesses

The paper does not substantiate it's goals with experiment. There are few recent citations and there is no detail into how the proposal contrasts with recent work, all citations are superficial. There are no substantial experiments or results.

Questions for authors

I do not have questions

1 comment

We respectfully disagree with the characterization of this work as a collection of lists or a superficial literature review. The central contribution is a theoretical framework for comparing mathematical theories of concepts through the operations they support, identifying the distinct capabilities and limitations that follow from their underlying mathematical commitments. This is a conceptual and theoretical paper, not an empirical ML paper; the categorization experiment is a case study illustrating the framework's implications, rather than its primary contribution. Nevertheless, it involves a comparative evaluation of nine mathematical frameworks against human judgments across three datasets. The claim that the paper presents no substantial experiments or results therefore seems to overlook a substantial part of the submission. More fundamentally, a theoretical contribution proposing a unifying perspective need not introduce a fully implemented mathematical system to be meaningful.

The review also makes broad claims about superficial citations and insufficient engagement with prior work, but provides no specific examples. We note that much of the relevant foundational literature is necessarily older, as the paper draws on seminal work in psychology and mathematics rather than restricting itself to recent ML publications. We would welcome specific examples of omissions or inaccuracies.

Finally, we are concerned by the unexplained ethics flag and the confidence rating of 4. The review identifies no ethical concern whatsoever, and we would appreciate clarification from the AC regarding the basis for this flag. We also find it difficult to reconcile the high confidence with the review's limited engagement with the paper's central theoretical contribution and its characterization of the empirical evaluation. We believe the submission deserves an assessment that engages with the actual nature and contributions of the work.

Reviewer_k8YhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a qualitative comparison of mathematical frameworks for concepts, based on how well they naturally support certain operations that one may want a concept representation to perform. The authors discuss thirteen such operations, and consider ten different frameworks, organized into three broad classes: content, relation, and process. The paper then presents experiments on categorization using nine of these frameworks, in order to validate their approach. Overall, I found the qualitative analysis to be interesting, but did not find the experiments to be convincing validations of the frameworks studied.

The main contributions of this paper are as follows:

  • A discussion of various operations that one might want to be able to perform using a concept representation (categorization, composition, revision, etc.), and a qualitative comparison of ten frameworks organized into three broad classes, in terms of how well they support these operations.
  • Experiments on categorization using nine of these frameworks (chosen because they can be implemented without supervision), in order to validate the above qualitative analysis.
Critical strengths
  • The focus on supported operations is a useful organizing principle that makes it easier to discuss the relative strengths and weaknesses of different approaches, compared to a less structured discussion of vector vs symbolic vs other representations. Also, the choice of operations considered is broad and includes some that go beyond categorization and similarity computation (e.g., composition, grounding, revision).

  • The analysis of agreement between frameworks is insightful. In particular, the discussion of the crowbar example nicely illustrates how agreement can result from a common error, and how this can lead to overestimating the validity of the corresponding approaches. This is a good point that would otherwise have been obscured by aggregate performance results.

Critical weaknesses
  • The notion of natural support is not well defined, and this makes the comparison between frameworks subjective. This is especially problematic given that this comparison is the main contribution of the paper. For example, Table 2 claims that probability provides natural support for grounding concepts, while for graphs grounding is not directly supported. However, in both cases it seems that some additional machinery would be needed (e.g., an observation model in the probabilistic case, or some mapping to data in the graph case). The distinction between what is canonical and what requires additional machinery seems subjective and this weakness was not overcome by the additional details in Appendix A.
  • The experiments do not provide convincing evidence that agreement with human categorization is explained by natural support of categorization in the mathematical framework, for the following reasons:
    1. Different frameworks are instantiated by models that differ not only in the underlying framework, but also in other respects (supervision, input representation, resources, coverage of items, etc.), so agreement with human categorization may be explained by any of these factors.
    2. Some of the connections between the models studied and the mathematical frameworks seem tenuous. For example, I do not see how the category-theory model based on Weisfeiler-Lehman refinement and clustering makes essential use of categorical structure, and how the dynamical model aggregating distances between representations at different checkpoints makes essential use of attractors.
Questions for authors
  • The categorization results you report may be due not just to the mathematical support but also to other factors that may vary across implementations, such as the amount/type of supervision, the input representation, the amount/type of data, algorithmic choices, etc. For example, grounding is supported by the observation model natively but only indirectly supported in the case of a graph and an external mapping. Can you comment on how your conclusions might vary if these factors were controlled for?

  • Relatedly, how robust are your conclusions to the particular implementations you chose for each framework? For instance, how representative is the Weisfeiler–Lehman implementation of GNNs? What category-theoretic structure is it instantiating? Similarly, how representative is the checkpoint-based implementation of attractor dynamics? Is there some evidence that this is actually exhibiting dynamics rather than an aggregation of static representations?

  • Table 2 proposes a categorization of different operations, but are there any specific predictions that can be made (and tested) using this table? For instance, beyond categorization, you could show how one prediction about composition/grounding/revision/etc. is useful.

1 comment

Thank you for the thoughtful review and for recognizing the value of the operation-based perspective. We would like to emphasize that the empirical results do provide evidence consistent with our account: frameworks that naturally support categorization through their canonical machinery (optimization and probability) achieve the strongest agreement with human judgments, followed by frameworks that support it through similarity-based structure (geometry and graphs), while frameworks requiring more substantial additional machinery perform less well. We agree that differences in resources and implementations prevent attributing these results exclusively to the mathematical frameworks, and do not claim a controlled causal comparison. Rather, the results illustrate how different mathematical commitments translate into different operationalizations and empirical outcomes. We will clarify this distinction, better explain the connections between the specific implementations and their underlying mathematical traditions, and sharpen the criteria distinguishing native from extended support. Finally, we will make more explicit how the framework generates predictions for operations beyond categorization.

Reviewer_mj2uOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an operation-based framework for studying concepts. It identifies thirteen conceptual operations, such as abstraction, categorization, composition, grounding, learning, similarity, and revision, and uses them to compare ten mathematical frameworks. The authors argue that different frameworks naturally support different operations depending on whether they treat concepts mainly as content, relations, or processes. A categorization case study shows that these mathematical commitments lead to different procedures and results.

Critical strengths
  • The paper addresses an ambitious and important question: whether concepts can be studied through a unified mathematical interface rather than through isolated representational traditions.
  • The operation-based framing is appealing. Evaluating formalisms by what operations they naturally support is a useful way to compare otherwise very different mathematical approaches.
  • The distinction between content, relational, and process frameworks is clear and helps organize a broad literature.
  • The categorization case study is a useful attempt to move beyond a purely conceptual taxonomy and test whether the proposed framework has empirical consequences.
  • The qualitative analysis of the “crowbar” example is insightful. It nicely shows that agreement across methods can reflect shared representational bias rather than independent evidence.
Critical weaknesses
  • The main contribution is primarily conceptual and taxonomic, but the criteria for assigning ✓, ◦, and × in Table 2 remain somewhat subjective. The paper provides rationales in the appendix, but many assignments are debatable. For example, probability is marked as not naturally supporting communication, while information theory is marked as naturally supporting it; logic is marked as not naturally supporting contextualization, although many logical extensions explicitly model context; and optimization is marked as not naturally supporting discrimination, even though discriminative learning is a central use of optimization in ML.
  • The thirteen operations are plausible, but the procedure used to derive them is not sufficiently formal or systematic. The paper says these operations recur across comparative cognition, cognitive psychology, and AI, but it does not clearly explain the selection process, inclusion/exclusion criteria, or whether important operations were omitted.
  • The empirical categorization experiment is interesting but limited as evidence for the broader thesis. It tests only one operation, categorization, and the operationalizations of the frameworks differ substantially in supervision, input representation, auxiliary resources, coverage, and evaluation metric.
  • The comparison between frameworks in Table 5 is difficult to interpret. Some methods are supervised and trained on labels, while others are unsupervised; some use embeddings, some use WordNet, and some use feature norms. As a result, performance may reflect data source or implementation choice more than the mathematical framework itself.
  • Several framework operationalizations seem too narrow to represent the corresponding mathematical traditions. For example, category theory is operationalized through a Weisfeiler–Lehman-style graph procedure, which may not capture the core strengths of categorical approaches. Similarly, algebra is represented through Formal Concept Analysis, which is only one algebraic instantiation.
  • The paper sometimes makes broad claims about what mathematical frameworks “naturally” support, but these claims risk oversimplifying rich literatures. Many hybrid or extended versions of these frameworks already exist, and the boundary between “native” and “added” machinery is not always clear.
  • The writing contains several grammatical and formatting issues, including phrases such as “offer but generalize poorly,” “information lose,” “comparisons is imperfect,” and “extrected.” These issues do not undermine the core idea but should be fixed.
Questions for authors
  1. How exactly were the thirteen conceptual operations selected? Was there a systematic literature review or coding procedure, or are they based on the authors’ synthesis?
  2. Can the criteria for ✓, ◦, and × be made more objective? For example, could the authors define a checklist for what counts as “canonical,” “induced,” or “added” support?
  3. How robust is Table 2 to alternative expert judgments? It would be helpful to report inter-rater agreement or some form of sensitivity analysis.
  4. In the categorization experiment, how can we separate the effect of the mathematical framework from the effect of the input resource, such as embeddings, WordNet, or feature norms?
  5. Why is category theory operationalized through WL-style structural equivalence? Are there alternative categorical constructions that might better reflect the framework’s compositional strengths?
  6. Can the empirical evaluation be extended to at least one additional operation, such as similarity, composition, or revision?
  7. Since the paper argues for hybrid formalisms, can the authors give a more concrete example of what such a hybrid mathematics of concepts would look like?
1 comment

Thank you for the thoughtful review and for recognizing the importance of the question and the value of the operation-based perspective. We would like to clarify that the paper does not claim to provide an exhaustive taxonomy of conceptual operations or a controlled comparison of mathematical frameworks under identical resources. Rather, it proposes an integrative perspective for examining what different formalisms make possible, using a representative set of operations and a case study to illustrate the consequences of different mathematical commitments. The differences in operationalization are therefore partly inherent to the traditions being compared, rather than simply confounds to be eliminated. We agree that the criteria for assessing operational support and the scope of individual instantiations could be articulated more explicitly, and will clarify these points. We hope the distinction between the paper's intended contribution and the stronger claims attributed to it addresses some of the concerns about its soundness.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.