CPADP: Adaptive Dropout Prediction Framework

• The paper introduces a staged dropout prediction framework that adapts its method from zero-shot LLM inference to fine-tuned PLM classification based on course progress.
• The framework leverages interaction logs over static demographics, achieving up to 95.4% accuracy in predicting dropouts in AI-driven courses.
• The system integrates prediction with intervention by generating personalized re-engagement emails to reduce dropout rates in Massive AI-empowered Courses.

Searching arXiv for the cited CPADP paper and closely related dropout-prediction work to ground the article. Course-Progress-Adaptive Dropout Prediction Framework (CPADP) is a course-progress-aware dropout prediction framework for LLM-driven interactive online courses. It was introduced in the study of Massive AI-empowered Courses (MAIC), where students interact with an AI Teacher, AI Teaching Assistants, and simulated peers in a text-based environment organized around slides and lecture notes rather than a predominantly video-based MOOC format (Wang et al., 24 Aug 2025). Within that setting, CPADP couples a log-derived definition of dropout with a staged prediction strategy that changes over the course timeline: zero-shot LLM inference at the beginning of a course, few-shot prompting when a small number of labeled cases become available, and fine-tuned pre-trained LLM (PLM) classification when sufficient task data has accumulated (Wang et al., 24 Aug 2025). The framework is closely tied to the broader problem of progress-sensitive dropout modeling, where longitudinal behavioral change, rather than static learner attributes alone, is treated as the primary predictive signal; a related example is the Dual-Modal Multiscale Sliding Window model, which likewise emphasizes time-varying risk and abrupt behavioral shifts in dropout prediction (Cheng et al., 16 May 2025).

1. Conceptual setting and definition of dropout

CPADP was proposed for Massive AI-empowered Courses, described as LLM-driven, interactive online courses that convert passive MOOCs into dynamic, text-based platforms. In the reported deployment, students learn by interacting with agents around slides and lecture notes, and each slide or chapter functions as a unit of interaction. The empirical study was conducted on the MAIC version of the course “Towards Artificial General Intelligence” (TAGI), spanning 6 chapters (Wang et al., 24 Aug 2025).

In this environment, dropout is defined from interaction logs rather than from enrollment records alone. Let $C$ be the set of all chapters in a MAIC course. A chapter $c \in C$ is “Study Complete” iff the teacher agent has presented all slides and lecture notes for that chapter. Let $S \subseteq C$ be the set of chapters a student has completed. The student’s course completion progress is defined as $P = |S|$. A student is considered to have dropped out if $S \neq C$, that is, if the student did not complete all chapters (Wang et al., 24 Aug 2025).

This definition is specific to MAIC and reflects the paper’s claim that dropout in interactive LLM-mediated courses should not simply be inherited from conventional MOOC formulations. A plausible implication is that CPADP is not merely a classifier attached to a pre-existing retention metric; it depends on a reconceptualization of completion that is native to chapter-wise agent interaction.

2. Empirical basis: what CPADP treats as predictive evidence

The MAIC study organizes its empirical analysis around three research questions: what might lead to dropouts, whether dropouts can be predicted, and whether dropouts can be reduced (Wang et al., 24 Aug 2025). The comparison among available information sources is central to the rationale for CPADP.

Basic information included college, major, gender, and grade. Chi-square tests showed no significant correlation with completion progress. Trait information was measured by questionnaire across Learning Motivation (LM), Academic Self-Efficacy (ASE), Persistence (LP), Strategy (SR), and Large Model Usage Frequency (LLMF), each rated on a 1–5 scale; Pearson correlations showed low correlation between these traits and completion progress (Wang et al., 24 Aug 2025).

The strongest relationship was found in interaction logs. The study reports that students with higher completion rates tend to have more frequent interactions, longer interactions, and more active or engaged textual exchanges. The paper further states that the model relies on interaction content, interaction frequency, interaction length, relevance of messages to course concepts, and whether interactions suggest active engagement or irrelevance or confusion (Wang et al., 24 Aug 2025). A case study contrasts “Fred,” whose irrelevant interactions were associated with high predicted dropout probability and actual dropout, with “Alice,” whose highly relevant, engaged interactions were associated with low predicted dropout probability and continued enrollment (Wang et al., 24 Aug 2025).

These findings position CPADP within a broader line of progress-sensitive dropout research. In a different educational context, the Dual-Modal Multiscale Sliding Window model also argues that abrupt changes in behavior, rather than static descriptors, act as key early signals of dropout risk (Cheng et al., 16 May 2025). The settings differ—MAIC is an online interactive LLM environment, whereas DMSW addresses offline K-12 records—but both studies converge on the view that temporal behavioral evidence is more informative than coarse demographic or trait variables.

3. Formal task specification and dataset construction

CPADP formulates dropout prediction as a supervised binary classification problem over a student’s historical interaction records. Given a student’s interaction records before the start of chapter $C_h$, denoted $\mathcal{I}_{C_h}$, the framework predicts whether the student will drop out from the start of chapter $C_h$ to the end of chapter $C_p$: $$\mathbb{P}(\text{Dropout} \mid \mathcal{I}_{C_h}, C_h, C_p), \quad C_h \in [1, \text{Last Chapter}], \quad C_p \in [C_h, \text{Last Chapter}]$$ Here, $c \in C$0 is the history boundary, $c \in C$1 is the end of the prediction period, and $c \in C$2 is the interaction-log prefix before chapter $c \in C$3 (Wang et al., 24 Aug 2025).

The authors explicitly describe this as course-progress-aware prediction. The prediction target changes with the student’s current place in the course, and the available evidence changes with the amount of observed interaction history. This differentiates CPADP from single-snapshot formulations that attempt to predict course withdrawal without conditioning on progress state.

For the TAGI course with 6 chapters, the dataset was built by varying $c \in C$4 and $c \in C$5. Each instance contains $c \in C$6, $c \in C$7, interaction records before $c \in C$8, and a binary label indicating whether a course withdrawal occurs. The reported dataset contains 1201 dropout prediction instances derived from the class records of 186 students, with a 20% test and 80% train split (Wang et al., 24 Aug 2025). The details note a discrepancy in an appendix table that mentions 198 students, but the main experimental section emphasizes 186 students (Wang et al., 24 Aug 2025).

This task design is structurally aligned with temporal dropout modeling more generally. DMSW likewise rejects a static endpoint-only view and instead tracks how student status evolves over time, using multiscale temporal windows to derive early warning features (Cheng et al., 16 May 2025). The two frameworks operationalize progress differently, but both treat dropout prediction as a longitudinal inference problem.

4. Architecture of the framework

CPADP consists of three stages, each matched to a different point in the course lifecycle (Wang et al., 24 Aug 2025).

1. Zero-shot stage

At the beginning of the course, when there are little or no labeled examples, CPADP uses a large pre-trained LLM to predict dropout from textual interaction history using prior knowledge and prompt instructions. No task-specific training data is required (Wang et al., 24 Aug 2025).

2. Few-shot stage

As the course progresses and a few labeled examples become available, prompt engineering is used and several labeled examples are inserted into the prompt. The study reports that this improves LLM inference quality over zero-shot prompting (Wang et al., 24 Aug 2025).

3. Fine-tuning stage

When enough data has accumulated, CPADP switches to a smaller pre-trained LLM. Features are extracted from interaction records and then passed into an MLP classifier for binary dropout prediction (Wang et al., 24 Aug 2025).

The PLM-based predictor is described formally as

$c \in C$9

followed by

$S \subseteq C$0

where $S \subseteq C$1 is the feature vector extracted from the interaction history (Wang et al., 24 Aug 2025).

The term “course-progress-adaptive” refers precisely to this stage-dependent switching of inference regime. Early in the course, the framework favors LLM prompting because labeled data are unavailable or scarce; later, it uses fine-tuning because the quantity of accumulated evidence makes supervised adaptation feasible. This suggests that CPADP is adaptive in two senses simultaneously: it is adaptive to learner progress and adaptive to the data regime induced by that progress.

5. Evaluation, prompt analysis, and predictive performance

The reported evaluation metrics are Precision, Recall, F1 score, and Accuracy, with results averaged over 3 runs with different random seeds (Wang et al., 24 Aug 2025). The comparative results are as follows.

Method	Setting	Accuracy
GPT-4	Zero-shot	0.716
GPT-4	Few-shot	0.779
GLM-4	Zero-shot	0.704
GLM-4	Few-shot	0.742
DeepSeek	Zero-shot	0.704
DeepSeek	Few-shot	0.662
PLM	Fine-tuning	0.954

The full table additionally reports that PLM fine-tuning achieved Precision $S \subseteq C$2, Recall $S \subseteq C$3, F1 $S \subseteq C$4, and Accuracy $S \subseteq C$5, which is the maximum accuracy reported for CPADP on the MAIC-TAGI dropout prediction dataset under the 80/20 train-test split (Wang et al., 24 Aug 2025).

The appendix also includes a prompt-design comparison for GPT-4. Zero-shot prompting achieved 0.716 accuracy; Random Selecting Examples achieved 0.730; Only False Examples achieved 0.745; Two Special Case Examples achieved 0.760; and Special Case Examples and Causal Examples achieved 0.779 (Wang et al., 24 Aug 2025). The factual takeaway is that few-shot prompt construction materially affected predictive quality, and the best prompting configuration was the one combining special cases and causal examples.

A further analysis presented as a heatmap over combinations of $S \subseteq C$6 and $S \subseteq C$7 shows that accuracy decreases as the gap $S \subseteq C$8 increases (Wang et al., 24 Aug 2025). The interpretation offered in the paper is that prediction becomes more difficult as the prediction horizon moves farther away from the available historical context. For deployment, the paper recommends using the student’s current chapter and prior chapters to assess the likelihood of completing the current chapter (Wang et al., 24 Aug 2025).

The contrast with DMSW is instructive. DMSW reports Accuracy $S \subseteq C$9, Precision $P = |S|$0, Recall $P = |S|$1, and F1-score $P = |S|$2, and describes its improvement primarily as an approximately 15-point gain in F1-score over its baselines (Cheng et al., 16 May 2025). The numerical results are not directly comparable because the settings, populations, labels, and modalities differ, but both papers report that explicit modeling of temporal progression yields better dropout-risk estimation than more static alternatives.

6. Intervention loop and deployment evidence

CPADP is not limited to risk estimation. The MAIC study extends the framework into a prediction-to-intervention pipeline by introducing a personalized email recall agent (Wang et al., 24 Aug 2025). Once at-risk students are identified using dropout probability predictions from CPADP, the system generates personalized reminder emails intended to re-engage them.

The email agent is prompt-based and uses each student’s interaction records, current course status, topics previously discussed, and likely interests or confusion points. The generated email may recall previous learning moments, reference content the student engaged with, use humor or inspiration, and nudge the student to return and continue learning (Wang et al., 24 Aug 2025). One example mentioned in the details is a student confused about “Hallucination” receiving a message that references the “General Artificial Intelligence Overview” session and previews relevant upcoming content (Wang et al., 24 Aug 2025).

The intervention experiment was conducted on January 10, day 65 of the semester. Predictions were made using a setting where $P = |S|$3 and $P = |S|$4 were both the chapters students were currently studying. Students with high predicted dropout probability were sent personalized emails (Wang et al., 24 Aug 2025).

The deployment evidence reported in the paper includes the following points. The appendix states that the MAIC system has been deployed for over one year, supporting dozens of courses, with more than 3,000 students (Wang et al., 24 Aug 2025). In the intervention analysis, there were 14 logins during Days 63–65 and 25 logins during Days 66–68, indicating increased engagement after the personalized emails were sent (Wang et al., 24 Aug 2025). Among the 17 students who logged in during Days 66–68, 9 were self-initiated and 8 were recalled by intervention. The self-initiated group had Offline Days 7.66, Chapter 2.56, Msg Num 8.78, and Msg Length 106.56; the recalled group had Offline Days 52.6, Chapter 0.75, Msg Num 0.25, and Msg Length 0.75 (Wang et al., 24 Aug 2025). The paper interprets this as evidence that the intervention brought back students who had been offline much longer and had minimal prior engagement.

This intervention component links CPADP to a broader educational objective: using progress-aware prediction to trigger timely support. A plausible implication is that the framework’s practical value depends not only on predictive discrimination but also on whether risk estimates can be translated into interventions that match the inferred source of disengagement.

7. Relation to adjacent dropout-prediction research and limitations

CPADP belongs to a larger family of dropout-prediction systems that emphasize longitudinal evidence, but its distinctive feature is the integration of course progress with language-model-based textual interaction analysis in an LLM-mediated classroom. The paper’s central empirical claim is that structured textual interaction logs are the most relevant factor for dropout behavior in MAIC, whereas basic demographic variables and measured traits exhibit weak or no meaningful association with completion progress (Wang et al., 24 Aug 2025).

Related work such as the Dual-Modal Multiscale Sliding Window model supports the general principle that dropout is often a process rather than a single event. DMSW argues that abrupt behavioral changes—such as sudden absenteeism, sharp academic decline, and severe punishments—act as key early warning signals, and it operationalizes that view through a multiscale temporal mechanism applied to textual behavioral records and numerical academic data (Cheng et al., 16 May 2025). The conceptual overlap suggests a common research direction: dropout prediction improves when models are designed around evolving trajectories rather than static learner profiles.

The limitations explicitly noted for CPADP are that the study uses one MAIC course, TAGI, so generalizability is limited, and that the intervention evaluation is relatively lightweight. The paper states that future work should include larger A/B tests, more courses, different email designs, and longer-term follow-up (Wang et al., 24 Aug 2025). Those caveats are significant because the framework’s empirical basis is tightly coupled to one interaction-intensive course and one chapter-based completion definition.

A common misconception would be to interpret CPADP as a general dropout predictor for all online education settings. The available evidence supports a narrower statement: CPADP was developed and evaluated for an LLM-driven interactive online course in which dropout is inferred from chapter completion within agent-mediated interaction logs (Wang et al., 24 Aug 2025). Another misconception would be to attribute its performance solely to large-model prompting. The reported experiments show that the best results came from the fine-tuned PLM plus MLP stage, not from zero-shot or few-shot LLM prompting (Wang et al., 24 Aug 2025).

In that sense, CPADP can be understood as a staged predictive framework whose defining idea is not merely “using LLMs for dropout prediction,” but aligning the prediction method with the amount of progress-dependent evidence available at a given point in the course.