Understanding the Role of Large Language Models in Personalizing and Scaffolding Strategies to Combat Academic Procrastination

The paper addresses academic procrastination—intentional task delay despite negative consequences—which is widespread among college students and detrimental to performance and wellbeing. Existing interventions (task completion, time management, motivational strategies) often fail to capture subjective, individualized contributors (e.g., learning styles, environmental and social contexts). LLMs could enable personalized scaffolding by processing open-ended, contextual inputs and engaging in dynamic interactions beyond static surveys or rule-based chatbots. Potential lies in contextualizing and integrating diverse procrastination management strategies, such as collaboratively creating micro-deadlines around a student’s schedule. However, LLMs lack direct perception of emotional states and complex social contexts; effectiveness depends on user-provided detail and informed prompt design. The study explores how students envision LLMs supporting task initiation, deadline management, emotion regulation, and social/environmental factors, while identifying limitations and guardrails. Research questions: RQ1: How do users envision LLMs tailoring strategies for managing academic procrastination? RQ2: What challenges or tensions might arise when using LLMs for tailoring strategies?

The related work covers two areas. (1) Application of LLMs in academic settings: LLMs can personalize learning and adapt to context across content creation, feedback, tutoring (e.g., debugging, hints), writing assistance, brainstorming, and microlearning. They offer dynamic, conversational support but raise ethical concerns, including overreliance, impaired critical thinking, diminished student-teacher interaction, and accuracy/hallucination risks—issues that are especially consequential for procrastination management where context and correctness matter. (2) Interventions to manage procrastination: Strategies focus on self-regulation (goal setting, time management), reflection and reminders (email/SMS nudges), technology-based interventions (blocking distractions, contextual cues), and social support. Prior systems show moderate benefits (e.g., Gantt Bot, reminders), but often fail to address individualized, nuanced factors. LLMs could fill this gap by handling open-ended inputs; however, user perceptions, integration with existing interventions, and ethical considerations need examination.

Design: A web-based technology probe (SPARK) using OpenAI GPT-4 generated context-specific advice for procrastination management. It included: (a) Seed messages grounded in psychological principles (cognitive insight, psychological flexibility, value alignment, implementation intentions); (b) Customization interface for open-ended situation input, tone (formal/informal), directedness (direct/indirect), length (50/100/150 words), inclusion of specific instructions, and optional custom prompts; (c) Features to edit LLM outputs and request iterative changes; (d) A future-self email drafting feature with optional LLM assistance and keyword scaffolding; (e) Instructions and examples to guide effective inputs. Participants: 15 university students (ages 18–25; mean 21.6±0.6; 10 women, 5 men; 8 Asian, 4 White, 2 African American, 1 Mixed-Race) residing in North America, recruited via email/social media; all self-identified as at least average procrastinators (IPS ≥ 24). To validate and contextualize findings, 6 experts (clinical psychology: 2; learning/education: 4; cognitive psychology: 2; 3 women, 3 men; 5 White, 1 Asian) were interviewed. Procedure: Semi-structured individual interviews (n=10; 30–50 minutes) and one focus group (n=5; 90 minutes) over Google Meet. Participants explored SPARK (about 10 minutes, with some requesting more time), typically: (1) viewing seed messages (0.5–2 minutes); (2) using customization options (6–8 minutes); (3) generating and refining personalized messages (2–4 minutes); (4) engaging in future-self email drafting (2–4 minutes). Participants then discussed utility, desired features, effort investment, guidance/resources, and routine integration. Experts later reviewed the themes from the student data and provided benefits/risks and ethical considerations. Analysis: Approximately 9.5 hours of audio produced 172 pages of transcripts, anonymized and cleaned. Two coders conducted thematic analysis with open coding on initial transcripts, consensus coding to refine a shared codebook, iterative application, and axial coding to group into themes. Nineteen final codes (e.g., need for direct solutions, step-by-step guidelines, use of calendars, clear disclaimers) were organized into themes. Ethics: Participants could skip questions or withdraw; interviewers were trained in Columbia-Suicide Risk Assessment protocols; no risks emerged during the study.

Themes and representative counts (participants P; experts E) from Table 2: (1) Aspirations for structured action steps: need for direct solutions (13P+2E); step-by-step guidelines (10P+4E); structuring recommendations (10P+3E); concerns about overreliance (2P+5E); desire for adaptive breakdown (3P+2E). Participants valued concrete, recipe-like steps; experts recommended detailing imminent steps without overwhelming users and warned against spoon-feeding that undermines independent thinking. (2) Deadline-driven instructions: reliance on deadlines (12P+4E); use of calendars (7P+2E); documenting goals (8P+5E); need for reminders (7P+4E); diversity in motivation (2P+4E). Participants wanted integration with calendars and work platforms (e.g., Google Calendar, Teams, Slack), early motivational cues, priority setting, and time estimates; experts suggested varied motivational strategies (e.g., WOOP, value-aligned affirmations). (3) Guided versus unguided questions: varied levels of guidance (6P+5E); multiple layers of features (7P+3E); flexible engagement (3P+2E). Some preferred guided sequences of open-ended prompts; others feared question overload becoming a new procrastination channel. Experts recommended adaptive questioning based on context (time, schedule, stress). (4) Concerns and boundaries for emotional support: need for emotional support (9P); limitations of LLMs in addressing emotions (1P+6E); clear disclaimers (5E). Participants wanted acknowledgment and resources; experts emphasized emotional validation with strong guardrails, transparency, and referral information for mental health support. (5) Support for using LLM tools: examples and instructions (5P+4E); collaborative elements (4P+2E); feedback to LLMs (5P+1E). Participants requested concrete examples, tutorials, social/collaborative prompt libraries, and mechanisms (e.g., star ratings) to give feedback and improve personalization. Overall, users prefer structured, deadline-centric, adaptable scaffolding with examples and guidance, while experts stress balancing support with autonomy and clear boundaries for emotional issues.

Findings address RQ1 by showing users envision LLMs as partners that convert vague, complex tasks into structured, stepwise, deadline-aware plans integrated with existing workflows, while offering examples and iterative customization. Users also want adaptive, context-sensitive questioning and collaborative resources to learn effective prompting and usage. For RQ2, tensions include risks of overreliance, cognitive overload from excessive options or probing, and ethical concerns about emotional/therapeutic guidance. The paper recommends design strategies: dynamic structured guidance that progressively reveals steps; integration with calendars and productivity platforms for prioritization and reminders; adaptive questioning tuned to user context (busyness, time of day); collaborative repositories of prompts and examples with feedback mechanisms; clear disclaimers and pathways to professional resources for emotional issues; and interventions that foster independent exploration and critical thinking to avoid learned dependence. These implications extend human-LLM collaboration literature by emphasizing balanced scaffolding, user autonomy, ethical transparency, and context-aware adaptivity.

The study demonstrates that LLM-based tools can help manage academic procrastination by providing structured, personalized, and deadline-focused guidance, supported by adaptive questioning and concrete examples. Using a GPT-4-powered probe (SPARK), interviews and focus groups with students and experts revealed strong preferences for step-by-step plans, calendar integration, reminders, and resources, alongside clear ethical boundaries for emotional support. The authors propose design recommendations for dynamic structured guidance, deadline-driven integrations, collaborative support features, flexible engagement modes, and prominent disclaimers with referrals. Future work should evaluate these features longitudinally in diverse populations and expand beyond academic procrastination to broader behavioral and productivity contexts.

Generalizability is limited by a North America-based sample and the focus on academic procrastination. The probe was not tested longitudinally without researcher oversight, limiting insights into long-term effectiveness and real-world adoption. Findings may not transfer to other behavioral or psychological domains. Future research should include more diverse, cross-cultural samples; ethically overseen longitudinal deployments; and exploration of LLM-based interventions across varied contexts (e.g., mental health, productivity, health behaviors).