Efficiency vs. Effectiveness: Self-Regulated Learning with LLM-Mediated Help-Seeking

Synthesis: Efficiency vs. effectiveness — a qualitative study of 20 STEM university students in Sweden showing that Large Language Models (LLMs) chatbots are woven into a layered, context-dependent Help-Seeking ecosystem rather than replacing human support. Students follow a four-stage process — deciding whether help is needed, choosing whom to ask, determining the type of help, and judging the help received — favoring instrumental over executive help-seeking to retain control of Problem Solving.

Key Findings

• A layered help-seeking ecosystem. LLMs do not displace peers and instructors; they become a low-barrier first step (e.g., "first ChatGPT, then classmates, and lastly teachers"). Peers remain important for conceptual negotiation and affective support, and instructors for complex, high-stakes issues.
• Four-stage LLM-mediated help-seeking process (adapted from Karabenick & Berger's model): (1) deciding whether help is needed — students try tasks independently first to preserve learning value; (2) choosing a source — ChatGPT as immediate first step, then peers, then instructors; (3) determining the type of help — hints, explanations, Scaffolding, routine-work streamlining, or extending learning; (4) judging the help — exercising selective trust, verifying AI outputs against coursework or with humans.
• Instrumental over executive help-seeking. Students favored instrumental help-seeking (hints, step-by-step guidance, concept explanations) over executive help-seeking (direct solutions), deliberately using ChatGPT as "a hint, an assisting tool, but not the standalone solution."
• Selective trust and verification. Students verify LLM outputs with peers, course materials, or instructors when accuracy is uncertain — e.g., "I only go to the TA if we can't tell whether ChatGPT is making things up."

What this means for practice

• Instructors. Teach verification and instrumental help-seeking explicitly: model seeking hints, explanations, and step-by-step guidance rather than direct solutions, since students who favored instrumental help kept control of Problem Solving.
• Instructors. Distinguish efficiency from offloading in task design — allow LLM use for syntax checks and quick clarifications, but protect independent work on debugging and cross-language programming, where shortcuts bypass the learning the task is meant to produce.
• Instructional designers. Build the four-stage help-seeking process (deciding whether help is needed, choosing a source, choosing the type of help, judging the help) into course structures so peers and instructors stay visible for conceptual negotiation and affective support.
• Researchers. Use the paper's draft survey items to measure SRL-for-LLM help-seeking, since the four-stage adaptation of Karabenick and Berger's model offers a framework for operationalizing an emerging construct.

Limitations

• The study interviewed 20 students (14 female, 6 male) at a single large Swedish STEM university, recruited by course announcements on Canvas, so volunteers likely over-represent students already using LLM chatbots.
• Data are 30–40 minute semi-structured interviews conducted between mid-May and July 2024 — self-reported accounts of study habits rather than observed help-seeking behavior or learning outcomes.
• There is no comparison or control condition and no performance measure; the findings describe students' stated intentions and reasoning, not what they actually did when studying.
• The findings sit in contrast to studies reporting over-reliance on LLM output, and the authors do not establish whether the difference stems from the population, the tasks, or the interview prompts used to elicit responses.

Citation

Viberg, O., Feldman-Maggor, Y., & Wong, J. (2026). Efficiency vs. Effectiveness: Self-Regulated Learning with LLM-Mediated Help-Seeking. Learning Letters, 8, Article 60.