From Literature to Research-Based Learning: An AI-Powered Information Extraction System to Enhance Undergraduate Thesis Completion

Synthesis: An and colleagues design and evaluate an AI-powered information extraction system that converts unstructured research publications into structured, traceable, and comparable datasets to support literature-driven comparison and undergraduate thesis completion in Higher Education. Grounded in Inquiry-Based Learning, cognitive load theory, and digital Scaffolding, the system is positioned as an epistemic scaffold that lets students inspect evidence–claim relationships through source traceability and structured comparison while reducing low-level data-handling demands. A mixed-methods pilot with 20 undergraduates and four supervisors suggests the system shifts students from isolated reading toward cross-study comparison and evidence-based justification, with implications for Self-Directed Learning, structured knowledge organization, and the responsible use of generative AI as a learning scaffold rather than a shortcut.

Key Findings

• High extraction coverage. Across the 80 uploaded documents, over 90% of the targeted experimental parameters and performance entries were successfully extracted and standardized (system log analysis).
• Reduced literature review time. Students' self-reported literature review time decreased by approximately 65%.
• Improved variable identification. Among the 16 junior students, the proportion able to identify at least three influential variables in experimental designs rose from 37.5% (6/16) to 87.5% (14/16) in a pre/post variable-identification prompt.
• Positive student perceptions. Over 90% of students agreed the system improved the efficiency and clarity of their literature review; ~85% said the structured tables and visual filters helped identify key experimental variables and performance trends; more than 80% found the interface intuitive.
• Reduced supervisory burden. All four supervisors reported a reduced need for repeated instruction and improved transparency of students' research logic, and 75% said the platform facilitated differentiated guidance by letting students present filtered data before consultation.
• Shift toward evidence-based reasoning. Students moved from passive reading toward cross-study comparison, trend-spotting, and methodological evaluation — behaviors associated with higher-order analysis and evaluation and growth in epistemic cognition within an Inquiry-Based Learning frame.
• Over-reliance caution. Interviews surfaced that normalized tables and comparison views can create a sense of completeness, tempting students to rely on extracted entries without checking boundary conditions; supervisors stressed treating the system as a structuring/indexing tool requiring explicit verification routines.

What this means for practice

• Instructors. Embed the extraction system in source-verification routines rather than releasing it as an answer source: normalized tables and comparison views can create a sense of completeness, so require students to check boundary conditions and verify the entries they intend to cite.
• Instructors. Use the traceable intermediate artifacts — extracted variables, comparison tables, source links, and workflow traces — to shift supervision from explaining basics toward formative, differentiated, data-informed guidance; all four supervisors reported less repeated instruction and 75% reported improved differentiated guidance.
• Instructors. Add scaffold-fading, reflective prompts, and check-the-source practice while the system offloads search, transcription, and unit standardization, so that epistemic Learner Agency survives the efficiency gain of roughly 65% self-reported literature review time saved.
• Designers. Recalibrate entity templates, field definitions, and domain ontologies, and validate them, before porting the pipeline to another STEM domain where parameter-based comparison is central.
• Researchers. Treat the pilot's quantitative results as descriptive process evidence rather than effects: with 20 undergraduates the study deliberately ran no inferential hypothesis testing.

Limitations

• Small, single-institution pilot with purposive recruitment: 20 undergraduates (16 junior, 4 senior) from four schools and four supervisors at one university, with quantitative results interpreted as descriptive and mechanism-oriented rather than statistically generalizable.
• Single-domain empirical validation: the pilot ran in materials-science thesis work, and full cross-domain porting had not been completed at the time of the study.
• Verification behavior was not quantified: although check-the-source was emphasized in onboarding and supported by traceability links, the study did not log how often students used the "Confirm" link or how many records each student cross-checked.
• Self-report and post-task perception data dominate: the roughly 65% reduction in review time is self-reported, most perception measures were collected post-task, and no assessment of unaided manual data-extraction skill tested whether the workflow transfers without the system.

Connected Concepts

• Inquiry-Based Learning
• STEM Education
• Higher Education
• Self-Directed Learning
• Knowledge Graph
• Generative AI

Connected Articles

• Fostering Self-Directed Growth with Generative AI: Toward a New Learning Analytics Framework — a conceptual framework treating GenAI as a non-prescriptive scaffold for self-directed growth, complementing this study's empirical view of AI as an epistemic scaffold
• Improving Capstone Team Outcomes through Dynamic Skill Matching and Preference Alignment — LLM-driven skill extraction and team formation in STEM capstone courses, another AI-for-higher-ed workflow
• Embodied Inquiry with AI as Facilitator: An Exploratory Case Study — where a language-based AI can stand within inquiry-based STEM learning, addressing a complementary boundary of AI as facilitator

Citation

From literature to research-based learning: An AI-powered information extraction system to enhance undergraduate thesis completion — An, R., Zhu, W., Zhao, Q., Laaksonen, A., & Lan, S. (2026). Computers and Education: Artificial Intelligence, 10, 100591.