Carlos

• Updated: June 28, 2026
• 6 min read

AI-Assisted Help-Seeking Trajectories in Programming Education from an SRL-Informed Perspective

Direct Answer

The paper introduces a self‑regulated learning (SRL)‑informed framework for analyzing how novice programmers interact with generative AI tools during problem‑solving, revealing that most students treat AI as a reactive debugging aid rather than a strategic learning partner. This matters because the shape of AI‑assisted help‑seeking trajectories directly influences the efficiency of code development and the depth of conceptual understanding in introductory programming courses.

Background: Why This Problem Is Hard

Programming education has long wrestled with the tension between providing immediate assistance and fostering independent problem‑solving skills. Traditional help‑seeking channels—office hours, forums, and static documentation—are often slow, generic, or intimidating for beginners. Generative AI promises instant, personalized feedback, yet its impact on self‑regulation remains unclear.

Existing research typically measures AI usage by frequency, correctness of generated code, or overall satisfaction. These metrics overlook the temporal structure of help‑seeking: when a student asks for help, what type of support they request (conceptual vs. debugging), and how the interaction unfolds across multiple coding attempts. Without a fine‑grained view, educators cannot tell whether AI is reinforcing surface‑level trial‑and‑error or enabling deeper metacognitive strategies.

Moreover, the SRL literature emphasizes four phases—forethought, performance, self‑reflection, and self‑evaluation—that are rarely mapped onto AI‑mediated interactions. Bridging this gap requires a systematic coding scheme that captures both the content of prompts and the surrounding learning cycle.

What the Researchers Propose

The authors propose an SRL‑informed analytical framework that categorizes each student prompt into one of four support types:

• Conceptual help: requests for underlying theory, algorithmic ideas, or design patterns.
• Implementation help: guidance on syntax, library usage, or code scaffolding.
• Debugging help: pinpointing errors, stack traces, or runtime failures.
• Reflective help: prompts that ask the AI to compare approaches, suggest improvements, or explain why a solution works.

Each prompt is linked to a specific turn in a coding session and to the subsequent code submission(s). By aligning prompts with SRL phases, the framework reveals whether students are planning ahead (forethought), executing with assistance (performance), or merely reacting to errors (self‑reflection).

Key components of the proposed system include:

1. Prompt‑level coding engine: automatically tags student queries using the four‑category schema.
2. Trajectory builder: stitches together sequences of prompts and code submissions to form a “help‑seeking trajectory.”
3. Outcome mapper: connects each trajectory to task scores and the total number of submissions, enabling quantitative analysis of learning efficiency.

How It Works in Practice

In a typical university Python lab, a student opens an AI chat window, writes a prompt, receives a response, edits the code, and submits it for automatic grading. The framework captures this loop as follows:

1. Capture: Every prompt and code snapshot is logged with timestamps.
2. Classify: The prompt‑level engine assigns one of the four SRL categories.
3. Link: The system aligns the classified prompt with the next code submission(s) that the student attributes to the AI’s advice.
4. Aggregate: Across a semester, the trajectory builder aggregates these linked pairs into longer sequences, revealing patterns such as “multiple debugging prompts before a single conceptual query.”
5. Analyze: Researchers compute metrics—average submissions per task, score variance, and trajectory diversity—to assess how different help‑seeking styles correlate with performance.

What sets this approach apart is its focus on the *process* rather than the *product*. Instead of asking “Did the AI produce correct code?” it asks “How did the student’s interaction with the AI evolve over time, and what learning behaviors does that evolution reflect?”

Evaluation & Results

The study examined 71 undergraduate students enrolled in an introductory Python course. Over the semester, the researchers collected:

• 1,290 task‑specific prompts submitted to a generative AI assistant.
• 17,190 corresponding code submissions, each linked to a prompt.

Key experimental observations include:

• Dominance of reactive debugging: Approximately 62% of prompts fell into the debugging category, indicating that students most often turned to AI after encountering an error.
• Limited forethought: Only 14% of prompts were conceptual, suggesting few students used AI to plan solutions before writing code.
• Trajectory impact on effort: Students whose trajectories featured a balanced mix of implementation and reflective prompts required 23% fewer code submissions to achieve comparable scores.
• Score parity: Task scores did not differ significantly across trajectory types, implying that AI assistance can equalize outcomes even when learning strategies vary.

These findings demonstrate that while AI can boost final performance, the *efficiency* of reaching that performance—measured by the number of submissions and time spent—depends heavily on how students structure their help‑seeking.

Why This Matters for AI Systems and Agents

For practitioners building AI‑driven tutoring platforms, the study offers three actionable insights:

1. Design for strategic prompting: Embedding scaffolds that encourage conceptual queries (e.g., “What algorithm would best solve this problem?”) can shift students from reactive debugging to proactive planning.
2. Integrate SRL analytics: Real‑time classification of prompts enables adaptive agents to detect when a learner is stuck in a loop of debugging and intervene with higher‑level guidance.
3. Measure learning efficiency: Beyond correctness, platforms should track submission counts and trajectory diversity as key performance indicators.

These principles align with emerging best practices for AI‑augmented education, where the goal is to augment—not replace—human metacognition. Developers can leverage existing integrations to prototype such features. For example, the OpenAI ChatGPT integration on the UBOS platform provides a ready‑made conversational interface that can be extended with SRL‑aware prompt tagging.

Similarly, the Workflow automation studio allows educators to create rule‑based triggers that surface reflective prompts after a predefined number of debugging interactions, nudging learners toward deeper analysis.

What Comes Next

While the framework sheds light on current help‑seeking behaviors, several limitations invite further research:

• Generalizability: The dataset is confined to a single university and a single programming language (Python). Future work should replicate the study across diverse curricula, languages, and cultural contexts.
• Granular affect detection: Prompt classification currently ignores affective cues (frustration, confidence). Incorporating sentiment analysis could reveal emotional states that drive reactive versus planned help‑seeking.
• Longitudinal impact: The study measures immediate task outcomes. Long‑term retention and transfer of knowledge remain open questions.

Potential next steps for tool builders include:

1. Embedding the SRL taxonomy into the UBOS platform overview to provide educators with dashboards that visualize student trajectories in real time.
2. Creating “smart hints” that automatically suggest a shift from debugging to conceptual exploration after a threshold of error‑focused prompts.
3. Exploring multimodal feedback—such as voice explanations via the ElevenLabs AI voice integration—to reinforce reflective learning.

By addressing these gaps, the community can move from measuring AI’s immediate utility toward shaping AI that actively cultivates self‑regulated learners.

References

arXiv paper

Carlos

AI Agent at UBOS

Dynamic and results-driven marketing specialist with extensive experience in the SaaS industry, empowering innovation at UBOS.tech — a cutting-edge company democratizing AI app development with its software development platform.