Learning Progress Monitoring Overview

• Learning Progress Monitoring (LPM) is a family of methods for tracking progress via longitudinal evidence, explicit targets, and adaptive feedback in varied educational and AI contexts.
• It spans LMS-centered analytics, competency tracking with CEFR-aligned scales, latent process modeling, and dynamic learner memory systems for personalized guidance.
• LPM enables practical insights by converting activity data into actionable signals while addressing challenges in fidelity, fairness, and model validation.

Learning Progress Monitoring (LPM) denotes the ongoing measurement and interpretation of movement toward a target of interest, but the literature uses the term across several technical settings. In LMS-centered e-learning, it is the systematic capture and analysis of learners’ activities within an integrated learning path to evaluate performance, support model-driven personalization, and foster trusted assimilation of knowledge (Alowayr et al., 2014). In competency tracking, it means ongoing, frequent measurement of specific language skills through CEFR-aligned outcomes and a consistent 5-point performance scale (Jr, 2013). In LP-guided science assessment, it links validated learning progressions, analytic rubrics, machine-learning scoring, and personalized feedback (Kaldaras et al., 16 Sep 2025). In latent-process modeling, it becomes distance-to-target inference in a shared metric space (Jeon et al., 2023). Recent AI work extends the term to dynamic learner modeling over daily LLM conversations (Cui et al., 26 Nov 2025), label-free monitoring of self-supervised representations (Xu et al., 2024), lifelong-learning evaluation (New et al., 2022), and noise-robust intrinsic motivation in reinforcement learning (Hou et al., 29 Sep 2025). Taken together, these uses suggest that LPM is best understood as a family of monitoring formalisms organized around longitudinal evidence, target specification, and feedback into adaptation.

1. LMS-centered monitoring and integrated learning paths

In the LMS literature, LPM is anchored in the claim that the Learning Management System is “the core of the entire e-learning process,” with monitoring implemented through web-server logs, LMS instrumentation, analysis tooling, and virtual environments (Alowayr et al., 2014). The reviewed workflow is high-level but explicit: LMS or web-server instrumentation logs all learner activity; MATEP ingests log files and related learner data; monitoring and analysis run at an administrator-defined update frequency; outputs are analyzed and visualized through pie charts and bar diagrams; and the resulting insights feed back into delivery through optimization of learning pathways and individualized support such as e-tutoring and e-mentoring. The tracked evidence includes course interactions and activity events written to logs, assessments and performance measurements captured by LMS tools and MATEP, and learner attributes drawn from academic and socio-demographic sources.

MATEP, the Monitoring and Analysis Tool for the E-learning Program, is the central monitoring tool in this line of work. It is described as a web application that runs over LMS web servers, uses log files to capture “every activity of learners,” updates automatically at an administration-defined frequency, and aggregates academic and socio-demographic information to provide instructors with a more accurate view of distance-learning activity and a basis for performance measurement (Alowayr et al., 2014). GISMO and CourseViz appear as monitoring and visualization tools used with LMSs, while eMUSE is associated with online knowledge sharing. Avatars-based virtual environments are treated as both an interaction modality and a monitoring context, because instructors can track learner activity in simulated settings.

The comparative survey reported in the same review, involving 150 participants, shows the relative visibility of these modalities. Technique usage was Avatars 50%, eMUSE 20%, GISMO 10%, and CourseViz 20%; preferred technology for effective e-learning was Avatars 44.5%, eMUSE 34%, CourseViz 16%, and GISMO 5.5%; and 80% affirmed that avatars frameworks deliver effective virtual environments for e-learning (Alowayr et al., 2014). At the same time, more than 50% reported that current MATEP tools do not fully meet requirements, and activity tracking remained problematic despite the available tools. The paper therefore presents LPM not as a solved instrumentation problem, but as a monitoring stack whose value depends on activity-capture fidelity, visualization, and the ability to translate traces into instructor action.

2. Competency tracking and progression-aligned assessment

A second major formulation of LPM treats progress as movement through explicit competency structures rather than as generic activity accumulation. In ESL/EFL monitoring, the proposed system uses Moodle outcomes or equivalent LCMS mechanisms to represent CEFR-aligned competencies, organized across six levels from A1 through C2, with roughly 30 grammar points per level and additional communicative functions defined through CEFR descriptors and rubrics (Jr, 2013). Teachers record evidence in the grader report through a 5-point performance scale: 5 = Mastery, 4 = Acceptable performance, 3 = Working performance, 2 = Limited performance, 1 = Minimal performance, while “-” is treated as a 1. A rating of 5 signifies mastery; programs may treat 4 as acceptable for progression, but full mastery is flagged at 5.

The implementation is explicitly longitudinal. Dedicated CEFR-level “courses” function as dossiers or gradebooks rather than instruction courses; learner records persist within those dossiers; pre-enrollment competencies can be merged into the current course; and records can be exported or transferred through Moodle’s backup/restore workflow (Jr, 2013). Monitoring is ongoing and manual: grammar evidence comes from question banks and tests aligned to grammar points, function evidence comes from in-person assessment using CEFR rubrics, and teachers update outcomes with “two or three clicks” after demonstration of performance. Because ratings are visible in learner and grader reports, the same data support student self-understanding, teacher planning, parent or supervisor oversight, and group-level curriculum decisions such as detecting B1 learners with A2 gaps or differentiating instruction between learners rated 1–3 and those already at 4–5.

LP-guided science assessment extends this competency logic to multi-modal constructs. In an NGSS-aligned high school Physical Science setting, student understanding of electrical interactions is represented through a validated learning progression whose broader structure includes levels 0–3, though the focal electroscope item operationalizes levels 0–2 (Kaldaras et al., 16 Sep 2025). Students submit both a drawing-based model and a short textual explanation. The analytic rubric contains 13 binary modeling categories and 8 binary explanation categories, combining accurate ideas and common inaccuracies. LP-level assignment is deterministic: for example, a level 2 model requires a score of 1 on at least 8 of categories 1–10, 0 on categories 11–13, and explicit evidence of more charge in B and stronger repulsion in B; a level 2 explanation requires category 16, the explicit causal link charge $\rightarrow$ force across scenarios, to equal 1, while categories 19–21 must be 0 (Kaldaras et al., 16 Sep 2025).

Machine learning is then used to automate this rubric-based monitoring. For drawings, the study uses a pretrained ResNet-18 with a modified final fully connected layer, feature dimension $d=512$, image size $224\times224$, Adam with learning rate $1\mathrm{e}{-4}$, and SMOTE augmentation; the dataset contains 1211 scored responses, split into 884 training and 327 testing cases (Kaldaras et al., 16 Sep 2025). For explanations, it uses BERT (bert-base-uncased) with dense layers, ReLU, sigmoid output, max sequence length 128, dropout 0.3, Adam with learning rate $2\mathrm{e}{-5}$, batch size 16, and early stopping on validation loss; the explanation dataset contains 1060 responses with an 80/20 split. Human scoring targeted Krippendorff’s alpha $\ge 0.8$ per analytic category. On the test set, core explanation categories tied to LP decisions reached acc 91.70% and F1 87.21 for C14, acc 97.56% and F1 89.36 for C15, and acc 92.19% and F1 80.95 for C16, while low-frequency categories such as C17 and C19 showed lower F1 values of 38.09 and 43.47 respectively (Kaldaras et al., 16 Sep 2025). This monitoring pipeline remains rule-based at the LP level: the models score rubric categories, and deterministic category combinations map those outputs to progression levels and personalized next-step feedback.

3. Analytical, probabilistic, and lifelong-learning formalisms

Another strand of LPM emphasizes formal inference over progress signals. In LMS analytics, one example is an advanced educational mining module integrated into an LMS to convert heterogeneous student behavior records into actionable knowledge (Zhang et al., 2024). The module standardizes data into categorical classes and discretized intervals, removes instances with missing variables, ensures no duplicates in each category, computes item probabilities $P_X = \mathrm{freq}(X_n^m)/N$, constructs candidate itemsets, and retains association rules under thresholds on support, confidence, and Conditional Probability Increment Ratio (CPIR). The paper gives the standard rule measures

$$\mathrm{Support}(X \Rightarrow Y)=\frac{\mathrm{freq}(X \Rightarrow Y)}{N},\qquad \mathrm{Confidence}(X \Rightarrow Y)=\frac{\mathrm{freq}(X \Rightarrow Y)}{\mathrm{freq}(X)}.$$

It also specifies $\theta_{\mathrm{Sup}}=0.7$ and $\theta_{\mathrm{Con}}=0.9$ for common rules, and support in $d=512$0, confidence $d=512$1, and $d=512$2 for exception rules (Zhang et al., 2024). On Dataset-I ($d=512$3, 28 attributes), runtime was 0.095 s for Faster Apriori versus 0.306 s for Apriori and 0.235 s for FP-Tree; on Dataset-II ($d=512$4, 28 attributes), runtime was 0.099 s versus 0.458 s and 0.177 s. The mined rules are then used for early warnings, dashboards, and intervention workflows.

A more explicitly statistical treatment appears in the latent process model for hard-to-measure targets. Individuals $d=512$5, items $d=512$6, and the target $d=512$7 are embedded in a common metric space, with the target defined as the centroid

$d=512$8

At baseline, responses depend on learner–item distance; for $d=512$9, they depend on learner–target distance. Progress is modeled through convex movement toward the target,

$224\times224$0

which yields the exact distance-reduction identity

$224\times224$1

The model is estimated in a Bayesian framework, using Polya–Gamma augmentation for binary data under a logit link and Metropolis–Hastings updates for latent positions and progress rates (Jeon et al., 2023). In the My Math Academy application, the dataset contains $224\times224$2 K–2 students, $224\times224$3 items, and two time points separated by about three months, and WAIC favored 2D or 3D embeddings over 1D and 4D.

Continual-learning evaluation provides a third formalism. The DARPA L2M framework organizes a lifetime into Learning Experiences (LX), Evaluation Experiences (EX), Learning Blocks (LB), and Evaluation Blocks (EB), and recommends clamping outliers, mapping normalized performance to $224\times224$4, and smoothing within LBs with a flat window of length $224\times224$5 (New et al., 2022). Transfer is defined through the contrast function

$224\times224$6

Using this formalism, Forward Transfer (FT), Backward Transfer (BT), Performance Maintenance (PM), Performance Relative to a Single-Task Expert (RP), and Sample Efficiency (SE) become canonical monitoring quantities. The practical blueprint built on this document also introduces Time-to-Competence, resource efficiency, stability–plasticity indices, and a composite lifetime score as LPM extensions rather than as metrics mandated by the original formalism (New et al., 2022). This suggests that, in lifelong settings, LPM is inseparable from scenario design: metrics are meaningful only when task introductions, revisitations, and evaluation events are explicitly structured.

4. Personalization, conversation traces, and dynamic learner memory

In mixed-initiative learning systems, LPM is increasingly tied to persistent learner models that are updated from everyday interactions. LOOM operationalizes progress toward mastery through module-level completions inside short personalized mini-courses grouped under broader goal umbrellas, with evidence drawn from explicit completion events, learner self-reports such as “already known” or “irrelevant,” and temporal activity signals derived from recent LLM conversations (Cui et al., 26 Nov 2025). The learner model is a lightweight hierarchy of goals $224\times224$7 courses $224\times224$8 modules, and conversations remain “active” for 10 days. This temporal activity filter is central to the continuity-versus-initiative balance: the system is responsive to current interests but preserves longer-term structure by regrouping courses into coherent goal umbrellas over time.

The architecture has four stages. First, daily chat transcripts are summarized into a one-sentence learner-centric statement, an umbrella theme, a difficulty tag, and an active flag. Second, recent summaries and the learner memory graph are used to propose 3–4 module mini-course outlines linked to source chats, with a succinct goal label, a course title, “questions you will answer,” and time estimates. Topic selection uses two modes: strengthen, which consolidates topics seen frequently in recent chats, and explore, which introduces adjacent concepts linked to the existing graph. Third, approved outlines are expanded into modules with concise lesson text and short multiple-choice quizzes, sequenced as core ideas $224\times224$9 applications $1\mathrm{e}{-4}$0 synthesis. Fourth, module completions, quiz attempts, and self-reports update module statuses and trigger regrouping operations such as merge, rename, or create goal umbrellas (Cui et al., 26 Nov 2025).

The description of LOOM also includes clearly marked extrapolated formalizations. These include a composite priority function

$1\mathrm{e}{-4}$1

where $1\mathrm{e}{-4}$2 is a recency score, $1\mathrm{e}{-4}$3 is a goal-alignment score, and $1\mathrm{e}{-4}$4 adapts to recent activity intensity, as well as a BKT-like mastery update in which $1\mathrm{e}{-4}$5 evolves through learning, forgetting, slip, and guess parameters (Cui et al., 26 Nov 2025). These equations are presented as principled extrapolations consistent with the system behavior rather than as equations specified in the prototype. Empirically, the formative study involved $1\mathrm{e}{-4}$6 participants over two days, each generating at least 10–15 new chats. Reported responses skewed positive for usefulness, coherence to prior chats, motivation to follow up, and willingness to reuse, while trust in correctness, length/depth matching time budgets, and repetition were more mixed. Participants also asked for stronger continuation and grouping, better filtering of instruction-focused sessions, more diverse formats, and optional quiz controls.

5. LPM inside machine-learning systems

Outside educational delivery systems, LPM also names mechanisms that monitor a model’s own learning dynamics. In reinforcement learning, the term denotes an intrinsic motivation scheme that rewards reductions in model prediction error rather than raw novelty or uncertainty, thereby addressing the noisy-TV problem (Hou et al., 29 Sep 2025). The dynamics model predicts $1\mathrm{e}{-4}$7, and the instantaneous error is the log mean-squared error

$1\mathrm{e}{-4}$8

A separate error model predicts expected previous-iteration error, and the intrinsic reward is

$1\mathrm{e}{-4}$9

The paper proves, under a Gaussian i.i.d. observation model with likelihood depending on log-MSE, that the intrinsic reward is zero-equivariant and a monotone indicator of information gain, and that the error model is necessary for the monotonicity correspondence (Hou et al., 29 Sep 2025). Empirically, LPM’s intrinsic reward converged to zero by about 150 steps in noisy MNIST, whereas AMA converged much later, around 400 steps; in a MiniWorld 3-room N-shaped maze with 2,176 visitable states, LPM achieved average coverage of 1,347.6 visited states, exceeding the second best by 95.3 states; and in Atari it was best on 4 of 6 games, with only a 3.9% drop on Space Invader, a 4.7% drop on UpNDown, and a 0.3% improvement on Ms PacMan under noise (Hou et al., 29 Sep 2025).

Self-supervised representation learning uses the term differently again. Label-free monitoring of SSL progress replaces labeled linear-probe evaluation with metrics defined directly on unlabeled embeddings: $2\mathrm{e}{-5}$0-means clustering, silhouette score, clustering agreement measured by adjusted mutual information, and entropy of the embedding distribution (Xu et al., 2024). At each milestone, every 20 epochs, the encoder produces 512-dimensional embeddings $2\mathrm{e}{-5}$1; UMAP with n_neighbors = 50 reduces them to $2\mathrm{e}{-5}$2; $2\mathrm{e}{-5}$3-means with cosine distance clusters $2\mathrm{e}{-5}$4 at $2\mathrm{e}{-5}$5 equal to the number of dataset classes and at $2\mathrm{e}{-5}$6; silhouette is computed on $2\mathrm{e}{-5}$7; and entropy is estimated from a 3D histogram with bin widths $2\mathrm{e}{-5}$8 for SSL models and $2\mathrm{e}{-5}$9 for ImageNet-pretrained transfers. The results are strongly method-dependent. AMI between the two clusterings correlates positively with linear-probe accuracy for SimCLR and MoCo-v2, but not for SimSiam: on CIFAR-10, the reported correlations are 0.71 and 0.91 for SimCLR with and without initialization, 0.64 and 0.80 for MoCo-v2, and 0.00 and $\ge 0.8$0 for SimSiam (Xu et al., 2024). Entropy generally decreases as learning progresses for SimCLR and MoCo-v2 after early instability, but increases for SimSiam; clustering-based metrics are said to be likely viable only for same-architecture comparisons, whereas entropy may be architecture-independent.

These AI-internal usages retain the central LPM intuition—monitor change in a target-relevant signal over time—but the monitored object is no longer a student. It is a learning system’s own progress, measured as error reduction, embedding organization, transfer, maintenance, or computational sustainability.

6. Trust, limitations, and open problems

Across the literature, LPM is repeatedly connected to trust, but trust is operationalized differently in different domains. In LMS-based e-learning, the core phrase is “trusted assimilation,” tied to learners’ need for assurance of content accuracy and authenticity, transparent activity capture, performance measurement, recognized specifications such as W3C-related attributes, and timely individualized support (Alowayr et al., 2014). In competency tracking, records are meant to be visible to educators, students, parents, and supervisors, yet privacy and security mechanisms are not specified; recommended practices include role-based access controls, informed consent for inter-institutional sharing, secure transfer of .mbz backups and spreadsheets, and data-retention rules (Jr, 2013). In LP-guided multi-modal science assessment, multi-modality is framed as equity support for culturally and linguistically diverse learners, but subgroup fairness analyses and calibration checks are not reported (Kaldaras et al., 16 Sep 2025).

Limitations are equally recurrent. The LMS review notes that activity tracking still falls short of stakeholder expectations, that MATEP would benefit from integrated data mining models, that SCORM and xAPI are not addressed, and that real-time alerting and algorithmic personalization specifications are not provided (Alowayr et al., 2014). The Faster Apriori LMS study depends on rigorous preprocessing, categorical or discretized inputs, and local validation before scaling; it also warns that sensitive attributes such as gender or family status can encode structural bias and require fairness-aware governance (Zhang et al., 2024). The latent process model leaves geometry and dimension selection as a substantive modeling decision, assumes time-invariant item positions and target, and does not detail a dedicated missing-data module (Jeon et al., 2023). LOOM is a formative, small-sample prototype with no baselines and a two-day study duration, and its prompting-based pipeline can yield variable granularity, factual inconsistencies, and overgeneralized mastery conclusions from single interactions (Cui et al., 26 Nov 2025). In reinforcement learning, the theoretical guarantees rely on an i.i.d. Gaussian observation model and accurate prediction of expected previous-iteration error by the auxiliary error model (Hou et al., 29 Sep 2025). In SSL, early-training outliers and UMAP sensitivity destabilize entropy estimates, and label-free clustering metrics can mislead for SimSiam or cross-architecture comparisons (Xu et al., 2024).

Future directions follow directly from these constraints. The literature explicitly calls for integrated data mining models and stronger activity-tracking fidelity in LMSs (Alowayr et al., 2014), automatic date stamping and expansion to additional skill areas in CEFR tracking (Jr, 2013), subgroup fairness analyses and possibly Rasch or Bayesian updating for longitudinal LP monitoring (Kaldaras et al., 16 Sep 2025), trend analytics and temporal modeling in mined LMS data (Zhang et al., 2024), explicit planning and mixed-initiative editing of learner graphs in conversational tutors (Cui et al., 26 Nov 2025), more robust entropy estimation and extension to additional SSL methods (Xu et al., 2024), and richer uncertainty-aware dynamics or episodic memory in RL exploration (Hou et al., 29 Sep 2025). Taken together, these directions suggest that no single metric, architecture, or statistical model exhausts LPM. Its common function is to convert longitudinal evidence into interpretable progress signals that can guide intervention, adaptation, or model selection, while remaining accountable to validity, trust, and domain-specific constraints.