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

The paper addresses how large language models (LLMs) could help manage academic procrastination, a prevalent issue among university students that negatively affects performance and wellbeing. Traditional interventions (e.g., time management tools, reminders, motivational strategies) often fail to capture individualized, subjective factors such as learning styles, emotional states, and environmental contexts. LLMs, with open-ended, conversational inputs and dynamic follow-ups, may tailor strategies to user-specific contexts by integrating a wide range of procrastination management techniques and offering context-aware advice (e.g., crafting micro-deadlines based on schedules and constraints). However, LLMs lack innate perception of users’ emotions and social contexts, requiring users to supply nuanced information, and there is a risk of misalignment with user needs or provision of non-validated guidance. The study explores two research questions: RQ1: How do users envision the role of LLMs in tailoring strategies for managing academic procrastination? RQ2: What challenges or tensions might arise when using LLMs for tailoring strategies to manage academic procrastination?

Related work covers two areas. 1) Application of LLMs in academic settings: LLMs can personalize and contextually adapt learning support, including writing assistance, code tutoring, content summarization, and personalized learning paths. They can scaffold complex tasks (brainstorming, structuring arguments, microlearning) but raise ethical concerns, including overreliance, reduced critical thinking, accuracy issues, and potential misinformation—especially problematic for procrastination management where erroneous advice can harm outcomes. 2) Interventions to manage procrastination: Prior approaches include self-regulatory skill development (goal setting, time management), reminders and reflective tools (emails, texts, distraction reflection), technological interventions (blocking distractions, resumption aids), and social support mechanisms. Many interventions struggle to address individualized, nuanced factors (social context, motivational triggers, lifestyle). LLMs’ open-ended inputs could help tailor interventions, but user expectations, interaction patterns, and ethical considerations remain underexplored.

The authors designed a web-based technology probe, SPARK (Self-generated Personalized Articulations and Reflections Kit), powered by GPT-4, to elicit participant feedback on LLM-enabled procrastination support rather than to validate a finalized system. SPARK components: 1) Seed Messages: Literature-grounded prompts incorporating four strategies—Cognitive Insight, Psychological Flexibility, Value Alignment, and Implementation Intentions—presented as templates for anti-procrastination guidance. 2) Interface options for LLM-generated, context-specific messages: Users provided open-ended descriptions of their situation and selected tone (formal/informal), directedness (direct/indirect), and length (50/100/150 words). They could opt to include concrete instructions and use predefined or custom prompts (e.g., coach tone, include a quote, add theory, keep concise, or user-defined). Users could request iterative changes or directly edit generated messages. 3) Prompting future action: A feature to draft an email to one’s future self, supported by LLM assistance and optional keywords (e.g., success, balance, independence) to anchor values and goals across time horizons (one week, three months, ten years), with the option to send the message. 4) Instructions and examples: In-situ guidance and examples for inputs and custom prompts aided users in effective interaction. Participants: 15 university students (ages 18–25; mean 21.6±0.6; 10 women, 5 men; diverse racial backgrounds) recruited in North America, self-identified as at least average procrastinators (IPS ≥ 24). Additionally, six experts (clinical psychology, learning/education, cognitive psychology; 3 women, 3 men) provided validation and critique. Procedure: Semi-structured interviews (n=10; 30–50 minutes) and one focus group (n=5; 90 minutes) were conducted via Google Meet. Participants explored SPARK for ~10 minutes (often longer), typically: reviewing seed messages; configuring interface options; generating and refining messages; and engaging with the future-self email feature. They then discussed utility, desired features, effort willing to invest, resources needed for personalization, and daily routine integration. All student participants received $20 USD/hour. Experts were later interviewed (45–60 minutes) to comment on themes and ethical considerations. Data analysis: Approximately 9.5 hours of audio yielded 172 transcribed pages (de-identified). Two coders conducted thematic analysis with open coding, consensus coding, iterative refinement, and axial coding. The final codebook comprised 19 codes (e.g., need for direct solutions, step-by-step guidelines, use of calendars, clear disclaimers). Ethical considerations: Participants could skip questions or withdraw; interviewers were trained in suicide risk assessment protocols, and referral procedures were prepared though not needed.

Five major themes emerged, supported by participant (P) and expert (E) mentions (counts from Table 2). 1) Aspirations for structured action steps: Participants preferred concrete, step-by-step plans and customization. Codes: need for direct solutions (13P+2E), step-by-step guidelines (10P+4E), structuring recommendations (10P+3E). Experts cautioned against overreliance (2P+5E), recommending a balance that supports independence and critical thinking. 2) Deadline-driven instructions: Users valued deadline-oriented planning and documentation. Codes: reliance on deadlines (12P+4E), documenting goals (8P+5E), need for reminders (7P+4E), use of calendars (7P+2E), adaptive breakdown (3P+2E), diversity in motivation (2P+4E). Participants suggested integration with calendars and communication tools (Google Calendar, Teams, Slack), priority setting, effort estimates, and early motivational prompts. Experts emphasized tailoring motivational strategies (e.g., WOOP, value alignment) and offering a repertoire of prompts. 3) Guided versus unguided questions: Preferences varied between sequenced, guided questions and minimal prompts to avoid overload or procrastination via questioning. Codes: varied levels of guidance (6P+5E), providing multiple layers of features (7P+3E), flexible engagement (3P+2E). Experts recommended adaptive questioning based on context (e.g., time of day, workload) and allowing short or long interactions. 4) Concerns and boundaries in providing emotional support: Participants desired emotional validation, resources, and optional pep talks; some sought support for more serious issues. Experts warned against LLMs offering therapy-like advice and urged clear boundaries. Codes: need for emotional support (9P), limitations of LLMs in addressing emotions (1P+6E), clear disclaimers (5E). 5) Providing support on the use of LLM-based tools: Users wanted examples, tutorials, and collaborative elements (e.g., shared prompts, social learning) and ways to give feedback (e.g., ratings). Codes: examples and instructions (5P+4E), collaborative elements (4P+2E), feedback to LLMs (5P+1E). Overall, participants preferred structured, deadline-focused planning with flexible guidance; experts emphasized ethical boundaries, adaptive interactions, and avoiding overreliance.

The findings address RQ1 by revealing that students envision LLMs as tools for producing structured, actionable, and deadline-focused plans that adapt to their context and workload, with options to customize tone, length, and specificity, and to receive examples and resources. They also value integrations with daily tools (calendars, messaging platforms) and collaborative learning mechanisms (shared prompts, ratings). For RQ2, tensions include risks of overreliance, reduced critical thinking, and the ethical boundary between supportive scaffolding and therapeutic advice. Preferences vary for guided vs. unguided questioning, suggesting adaptive, context-aware elicitation. Design implications include dynamic hierarchical task breakdown, integration with calendars and priority/effort estimation, multiple motivational framings, adaptive interaction flows, explicit disclaimers and resource referrals, and mechanisms for feedback and community sharing. These insights contribute to human-LLM collaboration literature by positioning LLMs as partners in task management and planning while preserving user autonomy and safety.

The study shows that LLMs can help personalize and scaffold strategies for managing academic procrastination by generating structured, deadline-oriented, and context-aware guidance. Through interviews and a technology probe with 15 students and 6 experts, the authors identify user desires for step-by-step planning, adaptive questioning based on busyness, integration with daily tools, and support in effectively using LLM features. Experts advise clear ethical boundaries, especially around emotional support, and caution against overreliance. The paper proposes design recommendations for dynamic structured guidance, calendar integration, diverse motivational prompts, collaborative repositories of prompts/responses, transparent disclaimers, and referrals to mental health resources. Future work should conduct longitudinal, real-world evaluations, engage more geographically and culturally diverse populations, and extend to other behavioral domains.

Key limitations include: 1) Sampling and context: Participants were North America-based university students; cultural and regional factors may limit generalizability. 2) Study design: SPARK was used as a technology probe under researcher oversight without longitudinal deployment; real-world effectiveness and sustained engagement remain untested. 3) Domain specificity: Findings center on academic procrastination and may not transfer directly to other behavioral or mental health domains. Future research should include diverse populations, ethically supervised longitudinal deployments, and exploration of broader applications (e.g., productivity, health behaviors).