CSDPM: Cognitive Structure Diffusion Model

• CSDPM is a discrete diffusion model that generates students’ cognitive structures as undirected graphs using educational priors and simulated interactions.
• It employs categorical graph diffusion, Graph Transformer denoising, and reinforcement learning with hierarchical SOLO rewards to refine latent concept and relation representations.
• Empirical results in Knowledge Tracing and Cognitive Diagnosis demonstrate faster inference and enhanced interpretability compared to traditional models.

Cognitive Structure Diffusion Probabilistic Model (CSDPM) most explicitly denotes a discrete diffusion model for generating students’ cognitive structures from educational priors and then aligning those generated structures with observed learning outcomes through policy optimization (Gu et al., 18 Aug 2025). In that formulation, a cognitive structure is modeled as an evolving undirected graph whose nodes represent a student’s construction of individual knowledge concepts and whose edges represent the student’s construction of inter-concept relations. The model combines categorical graph diffusion, Graph Transformer denoising, and reinforcement learning with SOLO-based hierarchical rewards. Related arXiv work also maps the label “cognitive structure diffusion probabilistic model” onto other cognitively motivated diffusion formalisms, notably predictive-coding diffusion and superstatistical dynamic cognitive modeling, but the explicit acronym CSDPM is introduced in the educational graph-generation setting (Gu et al., 18 Aug 2025).

1. Conceptual definition and graph representation

In the educational formulation, a student’s cognitive structure is defined as the subjective organization of an objective knowledge system, following the view that knowledge is psychologically organized through concepts and their relations (Gu et al., 18 Aug 2025). For student $s_i$ at timestamp $T$, the structure is represented as an undirected graph

$G_i^T = (V_i^T, E_i^T),$

with node tensor $V_i^T \in \mathbb{R}^{L \times c}$ and edge tensor $E_i^T \in \mathbb{R}^{L \times L \times c}$. Here, $L$ is the number of knowledge concepts and $c$ is the number of construction categories. Node and edge states are categorical over $\Delta^c$. Because the graph is undirected, the model operates on the upper-triangular edge tensor $E^+$ and then symmetrizes it.

A central difficulty is that ground-truth cognitive structures are not observable. The model therefore begins from simulated cognitive structures generated from educational priors derived from curriculum concept–item mappings and student interactions. Item-to-concept weights $\omega_{l,j}$ quantify how much question $T$0 assesses concept $T$1. From a student’s interaction history

$T$2

the paper defines rule-based functions for node and edge construction. The node construction function is

$T$3

and the edge construction function is

$T$4

These quantities lie in $T$5 and are treated as one-hot or small categorical distributions over construction levels. The resulting simulated structures embed educational priors without imposing an explicit prerequisite DAG. Instead, relational organization is induced through co-assessment weights and response patterns. This suggests that CSDPM is not merely a graph generator; it is a mechanism for converting partial pedagogical structure and behavioral traces into a latent representation of concept construction and inter-concept integration.

2. Discrete diffusion formulation

CSDPM uses a discrete categorical diffusion process rather than Gaussian diffusion (Gu et al., 18 Aug 2025). This modeling choice is motivated by the structured, sparse, multi-class nature of node and edge states. The forward noising chain factorizes over nodes and edges: $T$6

$T$7

For nodes, the categorical transition kernel is

$T$8

with

$T$9

where $G_i^T = (V_i^T, E_i^T),$0 decays from $G_i^T = (V_i^T, E_i^T),$1 to $G_i^T = (V_i^T, E_i^T),$2. When $G_i^T = (V_i^T, E_i^T),$3, there is no noise; as $G_i^T = (V_i^T, E_i^T),$4 decreases, the distribution is increasingly mixed with the uniform categorical distribution. Edge transitions are defined analogously.

The forward process admits closed forms: $G_i^T = (V_i^T, E_i^T),$5

$G_i^T = (V_i^T, E_i^T),$6

where $G_i^T = (V_i^T, E_i^T),$7. As $G_i^T = (V_i^T, E_i^T),$8 becomes sufficiently large and $G_i^T = (V_i^T, E_i^T),$9, node and edge states approach the uniform categorical distribution.

The reverse denoising model is factorized as

$V_i^T \in \mathbb{R}^{L \times c}$0

with

$V_i^T \in \mathbb{R}^{L \times c}$1

Its parameterization is of the $V_i^T \in \mathbb{R}^{L \times c}$2-type: the network predicts the clean node and edge states $V_i^T \in \mathbb{R}^{L \times c}$3 and $V_i^T \in \mathbb{R}^{L \times c}$4, which are then combined with the closed-form posteriors $V_i^T \in \mathbb{R}^{L \times c}$5 and $V_i^T \in \mathbb{R}^{L \times c}$6. For a node,

$V_i^T \in \mathbb{R}^{L \times c}$7

and the edge expression is analogous.

Pretraining maximizes

$V_i^T \in \mathbb{R}^{L \times c}$8

In implementation, this objective is realized with cross-entropy losses on nodes and edges: $V_i^T \in \mathbb{R}^{L \times c}$9

A common misconception is to treat all “diffusion probabilistic models” in cognition as Gaussian DDPMs. In the explicit educational CSDPM, the state space is discrete and categorical, and the diffusion kernels are categorical transition matrices rather than Gaussian perturbations.

3. Reverse-network parameterization and pretraining pipeline

The denoising network is an extended Graph Transformer that builds on graph transformers and discrete graph diffusion (Gu et al., 18 Aug 2025). Its inputs are the noisy graph $E_i^T \in \mathbb{R}^{L \times L \times c}$0, graph-theoretic features $E_i^T \in \mathbb{R}^{L \times L \times c}$1, and an embedding $E_i^T \in \mathbb{R}^{L \times L \times c}$2 of the current sample’s historical interaction sequence. The network predicts $E_i^T \in \mathbb{R}^{L \times L \times c}$3 and $E_i^T \in \mathbb{R}^{L \times L \times c}$4, the categorical distributions over clean node and edge states.

The pretraining pipeline proceeds in five steps. First, a simulated clean graph $E_i^T \in \mathbb{R}^{L \times L \times c}$5 and its interaction sequence $E_i^T \in \mathbb{R}^{L \times L \times c}$6 are sampled from the prior dataset $E_i^T \in \mathbb{R}^{L \times L \times c}$7. Second, a diffusion step $E_i^T \in \mathbb{R}^{L \times L \times c}$8 is sampled and a noisy graph $E_i^T \in \mathbb{R}^{L \times L \times c}$9 is drawn. Third, the model computes $L$0 and $L$1. Fourth, the Graph Transformer $L$2 outputs $L$3 and $L$4. Fifth, $L$5 is updated by cross-entropy to maximize $L$6, with node–edge balancing controlled by $L$7.

The reported implementation uses $L$8 diffusion steps in pretraining, an 8-layer Graph Transformer, AdamW, batch sizes in $L$9, dropout in $c$0, and learning rates in $c$1 with weight decay (Gu et al., 18 Aug 2025). The node–edge loss coefficient $c$2 is tuned, with best values around $c$3 for Math1, Math2, and FrcSub, and $c$4 for NIPS34. Initialization is Xavier. Reported hardware consists of Linux servers with Intel Xeon Platinum 8352V CPUs and NVIDIA RTX 4090 GPUs.

Inference times per generated structure are also reported: $c$5 ms for Math1 with $c$6 nodes, $c$7 ms for Math2 with $c$8 nodes, $c$9 ms for FrcSub with $\Delta^c$0 nodes, and $\Delta^c$1 ms for NIPS34 with $\Delta^c$2 nodes. These figures indicate that, although diffusion training is computationally heavier than typical KT or CD architectures, per-structure generation is relatively fast once the denoiser has been learned.

4. Policy optimization and SOLO-based hierarchical rewards

After pretraining, the generative process is optimized as a policy (Gu et al., 18 Aug 2025). The reverse diffusion is cast as a $\Delta^c$3-step Markov Decision Process in which each denoising step is an action that transforms $\Delta^c$4 into $\Delta^c$5. The mapping is: $\Delta^c$6

$\Delta^c$7

The transition dynamics are deterministic, and the only nonzero reward is terminal: $\Delta^c$8 The reinforcement objective is

$\Delta^c$9

The reward is defined through the SOLO taxonomy and evaluates how well the generated clean structure $E^+$0 aligns with the next observed interaction $E^+$1. The matching degrees are

$E^+$2

where $E^+$3 denotes logical XNOR, and $E^+$4 and $E^+$5 are the concept and relation subsets assessed by $E^+$6. The reward is piecewise: $E^+$7 with $E^+$8. The reward tuple used in experiments is

$E^+$9

To improve stability, the model adopts the eager policy gradient from GDPO: $\omega_{l,j}$0 The implementation normalizes rewards by subtracting the mean and dividing by the sample standard deviation. This policy stage is not an auxiliary component: it is the mechanism through which the diffusion sampler is aligned with “genuine cognitive development levels” rather than only with simulated prior structures.

5. Empirical performance in knowledge tracing and cognitive diagnosis

The model is evaluated on four real-world educational datasets: Math1 with $\omega_{l,j}$1 students, $\omega_{l,j}$2 exercises, $\omega_{l,j}$3 concepts, and $\omega_{l,j}$4 interactions; Math2 with $\omega_{l,j}$5 students, $\omega_{l,j}$6 exercises, $\omega_{l,j}$7 concepts, and $\omega_{l,j}$8 interactions; FrcSub with $\omega_{l,j}$9 students, $T$00 exercises, $T$01 concepts, and $T$02 interactions; and NIPS34, an EDNet subset, with $T$03 students, $T$04 exercises, $T$05 concepts, and $T$06 interactions (Gu et al., 18 Aug 2025). Interactions are split $T$07 at record level, with 5-fold cross-validation and significance testing by Wilcoxon rank-sum test.

The generated cognitive structures are used in downstream Knowledge Tracing (KT) and Cognitive Diagnosis (CD). KT baselines include DKT, SAKT, GKT, SKT, GRKT, MIKT, and ENAS-KT; CD baselines include IRT, MIRT, NCD, RCD, HyperCDM, and DisenGCD. Metrics are AUC, ACC, and RMSE.

For KT, CSG-KT reports the following best results. On Math1, AUC $T$08, ACC $T$09, RMSE $T$10; on Math2, AUC $T$11, ACC $T$12, RMSE $T$13; on FrcSub, AUC $T$14, ACC $T$15, RMSE $T$16; and on NIPS, AUC $T$17, ACC $T$18, RMSE $T$19. Each of these exceeds the cited KT baselines on the corresponding datasets.

For CD, CSG-CD likewise exceeds the reported baselines. On Math1, AUC $T$20, ACC $T$21, RMSE $T$22; on Math2, AUC $T$23, ACC $T$24, RMSE $T$25; on FrcSub, AUC $T$26, ACC $T$27, RMSE $T$28; and on NIPS, AUC $T$29, ACC $T$30, RMSE $T$31.

The ablation study separates the contribution of priors, diffusion pretraining, and RL alignment. Variant V1 uses simulation only; V2 adds diffusion pretraining; V3 uses RL with a generic reward but no pretraining; V4 uses RL with SOLO reward but no pretraining; V5 combines pretraining and generic reward; the full model combines pretraining with SOLO reward. Reported trends are monotone from V1 to V5 to full CSG across KT and CD, V1 often already matching or surpassing classical baselines, V3 generally outperforming V2, and both V4 $T$32 V3 and full CSG $T$33 V5 demonstrating the specific benefit of SOLO-based hierarchical reward shaping.

6. Interpretability, assumptions, and limitations

Interpretability is a central claim of the model because the generated cognitive structure is explicit at both concept and relation levels (Gu et al., 18 Aug 2025). The paper’s case studies show that incorrect predictions can arise from weak node construction, from weak edge construction despite strong nodes, or from both. In one example, student $T$34 before question $T$35 exhibits weak node and edge construction and is predicted incorrect; student $T$36 shows strong nodes but weak edges and is still predicted incorrect; student $T$37 shows strong node and edge construction and is predicted correct. Another example tracks student $T$38 across timestamps and depicts development from near-prestructural to fully constructed and integrated, qualitatively aligned with SOLO levels. The color coding distinguishes constructed, unconstructed, and low-frequency or irrelevant relations.

This explicit node–edge decomposition also reveals an important modeling assumption: cognitive development is represented as an undirected graph of categorical construction states. The paper does not enforce acyclicity, prerequisite consistency, or a directed knowledge graph. Relational coherence is instead learned from simulated priors and later refined by reward-based alignment. A plausible implication is that the model captures integration and co-construction more directly than prerequisite directionality.

Several limitations are stated. Diffusion models remain computationally heavier than standard KT and CD architectures. The rule-based simulation of cognitive structures simplifies complex and nonlinear cognitive development. No explicit prerequisite or DAG constraints are imposed. The authors identify future directions including fast samplers, consistency models, one-step diffusion, richer priors incorporating curriculum topology and prerequisite graphs, and differentiable penalties for structurally invalid graphs. These caveats are important because they delimit what the current CSDPM represents: not a directly observed psychological graph, but a learned generative approximation grounded in educational priors and aligned to subsequent student responses.

7. Related arXiv formulations and terminological scope

The acronym CSDPM is explicit in the educational graph model, but related arXiv work uses closely aligned language to denote other cognitively structured diffusion formalisms. One such mapping appears in CogDPM, where the paper does not define “CSDPM” directly but states that the mechanisms “exactly constitute a cognitive-structured diffusion probabilistic model” (Chen et al., 2024). There, the model connects diffusion probabilistic models with predictive coding: the unconditional reverse-step mean $T$39 plays the top-down generative role, the conditional reverse-step mean $T$40 plays the bottom-up perceptual role, and the residual

$T$41

is weighted by an inverse-precision estimate

$T$42

to yield the reverse update

$T$43

In that usage, “cognitive structure” refers to hierarchical residual correction and precision weighting in a DDPM-like reverse chain rather than to a graph of concepts and relations.

A second related formulation appears in neural superstatistics for dynamic cognitive models, where a “Cognitive Structure Diffusion Probabilistic Model” is presented as an instantiation of a hierarchical observation–transition model with a diffusion decision model as the local process (Schumacher et al., 2022). The general structure is

$T$44

with filtering posterior

$T$45

The local observation model is the diffusion decision model,

$T$46

with time-varying drift, threshold, starting point, and non-decision time; the transition model may be a Gaussian random walk or Gaussian process; and Bayesian inference is performed by an LSTM-based amortized neural estimator that learns $T$47 and $T$48. In that setting, “cognitive structure” refers to a superstatistical hierarchy of cognitive-state dynamics rather than to categorical graph generation or predictive-coding guidance.

These three uses establish that CSDPM is not a single universally standardized architecture. In the explicit educational usage, it is a discrete graph diffusion model with RL alignment (Gu et al., 18 Aug 2025). In related usage, it can denote predictive-coding diffusion with precision-weighted guidance (Chen et al., 2024) or a superstatistical diffusion-decision hierarchy with amortized Bayesian filtering (Schumacher et al., 2022). The common denominator is the combination of cognitively motivated structure with probabilistic diffusion-style generation or inference; the substantive object being modeled differs across formulations.