Cognitive Structure Generation (CSG)

• Cognitive Structure Generation (CSG) is a generative graph framework that models students’ subjective organization of knowledge by integrating concept and relation states.
• It leverages a cognitive structure diffusion probabilistic model with educational priors and reinforcement learning to align generated graphs with cognitive developmental levels.
• CSG enhances student modeling accuracy in knowledge tracing and cognitive diagnosis by providing interpretable, structured representations of learning processes.

Cognitive Structure Generation (CSG) is a framework for modeling and generating a student’s cognitive structure, where cognitive structure is defined as a student’s subjective organization of an objective knowledge system, reflected in the psychological construction of concepts and their relations. In the educational formulation, CSG addresses a long-standing challenge in student modeling and psychometrics by treating cognitive structure assessment as a generative graph problem: a Cognitive Structure Diffusion Probabilistic Model is first pretrained from educational priors, and its generative process is then optimized as a policy with hierarchical reward signals so that generated structures align with genuine cognitive development levels during learning (Gu et al., 18 Aug 2025). In a broader constructivist interpretation, this orientation is consistent with the claim that knowledge is actively constructed rather than passively copied, a position that also motivates schema-constrained generative memory architectures in agent systems (Zheng et al., 22 Apr 2026).

1. Definition and conceptual scope

Within the named CSG framework, cognitive structure is formalized as more than a scalar knowledge estimate or a flat concept profile. It is the student’s subjective organization of knowledge, and the task is to generate that organization from observed learning behavior rather than to assume it is directly measurable (Gu et al., 18 Aug 2025). This formulation is motivated by the observation that cognitive structure assessment has remained foundational yet largely unassessable in educational practice.

The framework is explicitly positioned against reductions that model only one side of the problem. In adaptive learning, existing methods often separately focus on either knowledge levels of learners or knowledge structure of learning items, even though both are important for learning path recommendation (Liu et al., 2019). CSG instead targets concepts and inter-concept relations jointly, making the generated object a structured latent state rather than a single mastery vector.

This emphasis also aligns with constructivist accounts of memory and knowledge. In schema-constrained agent memory, memory is fundamentally organized by cognitive schemas, and valid recall is treated as a generative process performed within schematic structures rather than as dense retrieval alone (Zheng et al., 22 Apr 2026). This suggests that, across domains, “cognitive structure generation” denotes not merely prediction from latent factors but explicit construction of structured internal organization.

2. Graph representation and educational priors

CSG models each student’s cognitive structure at timestamp $T$ as an evolving undirected graph

$$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$

where $\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$ represents node states for $L$ concepts and $c$ possible construction states per concept, and $$\mathcal{E}_i^T \in \mathbb{R}^{L \times L \times c}$$ represents relation states between concept pairs (Gu et al., 18 Aug 2025). The corresponding generation problem is the mapping

$f_{CSG}: X_i^T \rightarrow \mathcal{G}_i^T,$

with $X_i^T = \{(q_j, r_{ij})^t\}_{t=1}^T$ denoting the student’s answered questions and responses.

Because real cognitive structures are unobservable, the framework pretrains on simulated cognitive structures generated from educational theory-based rules. The node construction state for concept $k_l$ is written as

$\tilde{v}_{il}^T = f_{UOC}(k_l, X_i^T),$

and the edge construction state between concepts $$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$0 and $$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$1 as

$$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$2

with both functions defined from question-concept weights $$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$3, responses $$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$4, and interaction count $$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$5 (Gu et al., 18 Aug 2025). These simulated graphs form the dataset $$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$6 used for pretraining.

The representational consequence is substantial. Nodes encode the student’s construction state on each concept, while edges encode the student’s construction of inter-concept relations. The generated state is therefore intrinsically relational. This is precisely the dimension that prerequisite-graph methods in adaptive learning attempt to exploit at the content side: CSEAL represents learning items as nodes in a directed acyclic graph whose edges encode prerequisite relations, then uses cognitive navigation to constrain recommendation to logical candidates (Liu et al., 2019). CSG generalizes the idea from item structure alone to student-specific concept-and-relation structure.

3. Cognitive Structure Diffusion Probabilistic Model

The generative backbone of CSG is the Cognitive Structure Diffusion Probabilistic Model (CSDPM), a graph diffusion probabilistic model for discrete construction states (Gu et al., 18 Aug 2025). Its forward process corrupts simulated graphs into random noise through categorical transitions. For nodes,

$$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$7

with

$$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$8

Here $$\mathcal{G}_i^T = (\mathcal{V}_i^T, \mathcal{E}_i^T),$$9 decays from $\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$0 to $\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$1, interpolating between the original state and uniform noise.

The reverse process denoises the graph: $\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$2 For nodes, the model uses

$\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$3

Training maximizes a cross-entropy objective,

$\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$4

The reverse model is parameterized by an extended Graph Transformer with graph-theoretic features and learned embeddings for historical interactions as context (Gu et al., 18 Aug 2025).

The central technical claim is that CSG reformulates cognitive structure modeling as generative graph modeling. This is the mechanism by which an otherwise unobservable construct becomes operationalizable: not through direct annotation, but through prior-guided graph generation and denoising.

4. Reinforcement learning and SOLO-aligned policy optimization

After pretraining, CSG treats the denoising process as a $\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$5-step Markov Decision Process. The state is

$\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$6

the action is

$\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$7

and the policy is

$\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$8

Rewards are zero except at the final step, where the generated graph is evaluated against hierarchical developmental criteria (Gu et al., 18 Aug 2025).

Those criteria come from the SOLO taxonomy: Prestructural, Unistructural, Multistructural, Relational, and Extended abstract. Given generated $\mathcal{V}_i^T \in \mathbb{R}^{L \times c}$9, next test item $L$0, and response $L$1, CSG computes node-level matching

$L$2

and edge-level matching

$L$3

where $L$4 is the XNOR operator (Gu et al., 18 Aug 2025). A piecewise reward function $L$5 then assigns $L$6 according to the joint conditions on $L$7 and $L$8.

The reinforcement-learning objective is

$L$9

with policy gradient

$c$0

The role of this stage is not generic fine-tuning; it is explicit alignment of graph generation with cognitive developmental levels during learning (Gu et al., 18 Aug 2025).

5. Downstream student modeling, empirical performance, and interpretability

Generated cognitive structures are used as student state representations in both knowledge tracing (KT) and cognitive diagnosis (CD). For KT,

$c$1

and for CD,

$c$2

where Pooling is an edge-aware hard clustering graph pooling, $c$3 is element-wise multiplication, FC is a fully connected layer, and $c$4 is the sigmoid function (Gu et al., 18 Aug 2025).

On four real-world datasets, the reported AUC values for the CSG-based models are as follows (Gu et al., 18 Aug 2025):

Dataset	CSG-KT AUC	CSG-CD AUC
Math1	0.8192	0.8104
Math2	0.7813	0.8193
FrcSub	0.8575	0.8757
NIPS34	0.7339	0.8002

In the reported comparisons, CSG-KT outperformed DKT, SAKT, GKT, SKT, GRKT, MIKT, and ENAS-KT, while CSG-CD outperformed IRT, MIRT, NCD, RCD, HyperCDM, and DisenGCD across the listed datasets (Gu et al., 18 Aug 2025). The broader claim made for the framework is that generated cognitive structures provide more comprehensive and effective representations for student modeling, substantially improving performance on KT and CD tasks while enhancing interpretability.

Interpretability is a first-class property of the framework rather than a post hoc explanation layer. The generated graphs expose nodes as student mastery on each concept and edges as construction of inter-concept relations. Reported case studies show graph evolution from sparse structures associated with prestructural states toward fully connected or broadly mastered structures associated with extended abstract states (Gu et al., 18 Aug 2025). This makes it possible to inspect whether an incorrect answer is attributable to weak concept construction, weak relation construction, or both.

6. Related formulations of cognitive structure across AI and cognitive science

CSG sits within a larger family of approaches that operationalize internal structure rather than flat latent state. In adaptive learning, CSEAL combines recurrent knowledge tracing, a prerequisite DAG, cognitive navigation, and an actor-critic recommender. Its stated contribution is to exploit the multifaceted cognitive structure for learning path recommendation, and its experiments report higher effectiveness and logicality than methods that model only one component of the structure (Liu et al., 2019).

In agent memory, SCG-MEM replaces dense retrieval with Schema-Constrained Generation. It maintains a dynamic Cognitive Schema as a Prefix Trie, uses schema-constrained decoding to ensure that only valid memory keys can be generated, models updates through assimilation and accommodation, and overlays an Associative Graph for multi-hop reasoning. On LoCoMo, it reports improvements across all categories over retrieval-based baselines and provides a formal guarantee against structural hallucinations (Zheng et al., 22 Apr 2026).

In social creativity, cognitive structures are semantic networks $c$5 that share a vocabulary but differ in topology. A single rewiring probability $c$6 in a Watts–Strogatz process tunes modularity, and a random walk on the resulting graph produces ideational breadth $c$7. The reported empirical result is a strong negative relationship between semantic modularity and ideational breadth, $c$8, and at the collective level lower pre-interaction overlap predicts larger stimulation gains while shared inspiration sources induce redundancy (Ahmed et al., 3 Feb 2026).

In language neuroscience, structure-building is operationalized with Combinatory Categorial Grammar parsing rather than as generic next-word predictability. CCG-derived measures reportedly provide a superior fit to neural signals compared to CFG-derived measures, predominantly in the left posterior temporal lobe, while predictability effects are spatially distinct in bilateral superior temporal cortex (Stanojević et al., 2022). This suggests that structure-building and predictability are separable processes.

In emergent-cognition architectures, COGENT3 treats computational structures as dynamically emergent from agent interactions in a triadic hypergraph, with temperature modulation, memory kernels, and pattern formation networks regulating formation and dissolution of functional organization (Salazar, 5 Apr 2025). Across these domains, the recurring theme is that cognition is modeled through explicit structure construction, not only through vector similarity, fixed module assignment, or outcome prediction.

7. Terminological overlap and disambiguation

The acronym CSG is not unique to cognitive structure generation. In geometry processing and CAD, it commonly denotes Constructive Solid Geometry, where shapes are represented as trees or fixed-depth Boolean assemblies of primitives. Representative lines include CSG-Stump, a three-level complement–intersection–union representation equivalent to CSG, D$c$9CSG for unsupervised learning of compact CSG trees with dual complements and dropouts, and surveys of methods for converting unstructured data such as point clouds and polygon meshes into CSG models (Ren et al., 2021, Yu et al., 2023, Fayolle et al., 2023).

In multiagent systems, CSG also denotes Coalition Structure Generation, the problem of partitioning agents into exhaustive and disjoint coalitions to maximize social welfare. This literature includes GRASP with path-relinking, the hierarchical dynamic-programming algorithm DyPE, the large-scale anytime search algorithm SALDAE, and sparse probabilistic formulations that learn coalition values from episodic data before solving the combinatorial optimization problem [(Mauro et al., 2011); (Vinyals et al., 2013); (Taguelmimt et al., 14 Feb 2025); (Majumdar, 1 Jan 2026)].

This terminological overlap makes the expanded form important in cross-disciplinary contexts. Within educational AI and adjacent constructivist work, however, Cognitive Structure Generation now denotes a specific generative-graph framework for representing concepts and inter-concept relations, pretrained from educational priors and optimized against developmental reward signals (Gu et al., 18 Aug 2025).