Learner-State Features in Adaptive Learning

• Learner-state features are structured variables that capture a learner's evolving cognitive, affective, and behavioral states via objectives and motivations.
• They are integrated through methodologies like Pxplore, CSCD, and LSKT to support personalized assessment and reinforcement learning policy design.
• Empirical evidence shows enhanced alignment rates and predictive performance when multimodal data such as language, video, and interaction logs are fused.

Learner-state features are structured variables and representations designed to capture, evaluate, and utilize a learner’s evolving state within educational, cognitive, or autonomous systems. These features can encode objectives, mastery, knowledge structure, affective signals, behavioral patterns, and latent readiness, serving as foundations for personalization, assessment, and decision-making in adaptive learning, cognitive diagnosis, and RL agent design. The formalization, extraction, and integration of learner-state features is a central methodological axis in contemporary research, shaping the capabilities of systems ranging from personalized path planners to large-scale knowledge tracing models and autonomous decision-making agents.

1. Formal Definitions and Structural Models of Learner-State Features

Recent frameworks model learner-state features as high-dimensional, structured records. Pxplore (Lim et al., 15 Oct 2025) introduces a four-set tuple at time $t$: $s_t = (O^L_t, O^S_t, M^I_t, M^E_t)$ where $O^L_t$ is the set of long-term objectives, $O^S_t$ short-term objectives, $M^I_t$ implicit motivations inferred from behavior, and $M^E_t$ explicit motivations stated by the learner. Each component $c$ has five fields: $$c = \{\mathrm{description},\,\mathrm{metric},\,\mathrm{evidence},\,\mathrm{confidence},\,\mathrm{status}\}$$ Status is operationalized as $[\mathrm{ALIGNED}]$ or $[\mathrm{NOT\_ALIGNED}]$ based on the metric threshold.

In cognitive diagnosis (CSCD (Chen et al., 2024)), learner-state features are the fusion of Knowledge State (KS)—mastery scores per concept—and Knowledge Structure State (KUS)—embeddings of dependency and predecessor–successor relations among concepts refined through edge-featureted graph attention networks (EGAT).

For knowledge tracing, LSKT (Wang et al., 2024) distinguishes instantaneous latent “learning state” $\hat y_t$ (short-term readiness or cognitive state) from “knowledge state” $h_t$ (aggregated mastery), extracting $\hat y_t$ by causal convolution over sequential historical interactions.

Tabular comparison of core learner-state feature models:

Framework	State Structure	Component Types
Pxplore	4-set tuple: $O^L$, $O^S$, $M^I$, $M^E$	Objectives, Motivations (explicit, implicit)
CSCD	KS, KUS fused concept/edge GAT embeddings	Concept mastery, structural relations
LSKT	$\hat y_t$, $h_t$ (conv. state, knowledge)	Learning state, knowledge state, embeddings

2. Extraction and Construction of Learner-State Features

Feature extraction spans manual, automated, and generative protocols. In Pxplore (Lim et al., 15 Oct 2025), the function $F_{\mathrm{eval}}$ uses LLM-based evaluators to parse dialog evidence $I_t$ and the prior state $s_t$, updating components by checking metrics (e.g., understanding scores, question types), flipping $[\mathrm{ALIGNED}]$ status, appending new objectives, and recalculating confidence.

Multi-channel behavioral models (Tian et al., 2018) extract features from synchronous video (deep-CNN facial, head-pose angles), eye tracking (statistical, wavelet, Fourier descriptors), and mouse dynamics. Early fusion (feature-level PCA) is preferred for precision, with time-synced vectors constructed for supervised regression.

Language-based discrimination of confusion (Atapattu et al., 2019) employs lexical, syntactic, sentiment/emotion, and discourse features—type-token ratio, post length, pronoun profiles, sentiment markers, negation, question bigrams—preselected by MANOVA and model importance, yielding a cross-domain predictor of individual affective states.

In RL and agent state modeling (Michlo et al., 2022, Wang et al., 2024), state-representation features are constructed by LLM code generation (LESR), triplet-based metric learning (VAE+triplet), or generate-and-test feature construction (linear traces, Boolean imprints in (Samani et al., 2021)).

3. Mathematical Formulation and Dimensionality

The formal dimensionality of learner-state features reflects both compositional hierarchy and temporal dependence.

• Pxplore’s variable-dimension $s_t$ scales with the sum $n_L + n_S + n_I + n_E$ (typical session ~180 components), each with 5 fields.
• CSCD employs multi-head edge-featured GATs to generate $\mathbb{R}^d$ embeddings for KS and KUS, fusing them via learned within- and across-relation attention scores.
• LSKT’s learning state $\hat y_t \in \mathbb{R}^D$ arises from causal convolution over interaction embeddings, while historical and situational context features in LKT (Jr. et al., 2020) are computed by precise formulas from timestamped practice logs ($$n_{s,k,t}, \mathrm{age}_{s,k,t}, \mathrm{spacing}_{s,k,t}$$, etc.).

In RL, the feature vector is built by concatenation: for LESR (Wang et al., 2024),

$$s^\mathrm{c} := \phi(s) = [s; F(s)] \in \mathbb{R}^{d+p}$$

and triplet-regularized latent codes (Michlo et al., 2022) via VAE encoder mean outputs.

4. Integration with Reward, Prediction, and Policy Learning

Learner-state features fundamentally shape reward functions, inference, and decision policies.

• Pxplore’s reward for action $k_{t+1}$ is defined by state-component alignment: $$R(s_t, k_{t+1}) = \sum_{c \in C(s_{t+1})} w_c\,\mathrm{conf}(s_{t+1},c)\, [\mathbb{I}(s_{t+1},c) - \mathbb{I}(s_t,c)]$$ Only newly achieved alignments contribute.
• Markov Decision Process: policy $\pi^\ast$ maps $s_t \to \Delta(K)$, trained by SFT and GRPO using group-normalized advantages.
• In CSCD (Chen et al., 2024), state features are fused for diagnostic prediction through a MIRT-style interaction layer: $$x = Q_e \circ (h^s - h^\mathrm{diff}) \times h^\mathrm{disc}$$ Binary cross-entropy loss guides training.

LSKT (Wang et al., 2024) fuses learning and knowledge states: $z_t = [h_t\,\Vert\,\hat y_t]\,W_7$ and predicts future response probability conditioned on both, with attention distributions $\gamma_{t,\tau}$ reflecting cluster-aware similarity.

Language-based confusion features (Atapattu et al., 2019) are concatenated as real-valued vectors for classification, with Random Forest ranking their predictive value across domains.

5. Empirical Findings and Per-Feature Impact

Empirical evaluations disentangle the contribution of specific learner-state features.

Pxplore (Lim et al., 15 Oct 2025)’s Table 3 reports alignment rates for each dimension:

• OL (Long-term objectives) up to ≈ 67%
• OS (Short-term objectives) up to ≈ 64%
• MI (Implicit motivations) up to ≈ 80%
• ME (Explicit motivations) up to ≈ 67%

SFT and GRPO both improve alignment rates, with explicit motivations being the hardest to align via prompt-only LLM inference.

CSCD (Chen et al., 2024) demonstrates through ablations that diagnostic accuracy is maximized only when both KS and KUS are modeled jointly; models omitting one component suffer in accuracy and interpretability.

LSKT (Wang et al., 2024) shows that incorporating learning state, especially with fine-grained IRT-derived embeddings (3PL), improves AUC by up to 3.33% on large benchmarks, with ablations confirming that both knowledge and learning state are synergistically complementary.

Multimodal fusion (Tian et al., 2018) yields R² ≈ 0.98 with all channels (video, eye, mouse), outperforming single-channel approaches (eye alone R² ≈ 0.43).

Language discourse features (Atapattu et al., 2019) such as TTR, negation, pronoun usage, and sentiment indices are among the strongest individual predictors of confusion across education, humanities, and medicine, with cross-domain F₁ scores exceeding 70%.

6. Adaptation, Generalization, and Domain Extension

The formalism and extraction methodology of learner-state features supports adaptation to new domains and states.

• Feature sets derived from language, behavior, or latent signals can be extended by incorporating new lexicons (frustration, boredom), temporal thread context, or dialog-acts.
• Structural models (Pxplore, CSCD, LSKT) support the addition of new objective types, motivational components, or context features without re-architecting core logic.
• Multimodal pipelines (Malekshahi et al., 2024) are designed to flexibly integrate new modalities (e.g., gaze, action units) and utilize adaptation policies for label mapping across affective dimensions.

In logistic knowledge tracing (Jr. et al., 2020), feature-based intercepts (propdecStudent, logitdecStudent) replace fixed student parameters to enhance generalizability to unseen learners.

7. Theoretical Frameworks and Representational Properties

Underlying learner-state representations are combinatorial learning spaces (0803.4030) (antimatroids, lower sets of posets), Bayesian state inference, and algebraic decompositions. Learning sequences provide the foundational enumeration and indexing for adaptive systems such as ALEKS.

Learning state extraction by convolution, metric learning with triplet losses, feature construction via generate-and-test, and LLM-driven code generation for RL agents reflect the technical diversity in constructing meaningful, generalizable, and procedurally actionable learner-state features.

The impact of representation smoothness, as evidenced by empirical Lipschitz continuity reduction (LESR (Wang et al., 2024)), figures centrally in theoretical guarantees of policy improvement and sample efficiency.

Overall, learner-state features constitute a multidimensional, context-sensitive, and computationally tractable substrate for individualized learning, cognitive assessment, and autonomous agent optimization, as advanced in the recent literature.