Where should an AI tutor’s knowledge of a student’s interests come from?

Interest-based learning connects instruction to topics each student cares about. Large language model (LLM) tutors could scale it if they identified students’ interests automatically and effectively. My dissertation asks, as its abstract puts it: “From what should student interests be identified?”

A tutor teaches from what reaches it

A tutor teaches not from the interest itself but from the information about it that reaches the tutor. So I compared where that information comes from and measured what each source delivers.

1. Where interests come from

   • Asking the studentStudy 1
   • Browsing historyStudy 2
   • Social-media postsStudy 2

2. What reaches the tutor

   • Interest information unitscoded as affect, function, knowledge or engagement

3. How I judge identification

   • Accuracy
   • Depthnumber of units
   • Richnessnumber of categories covered
   • Phasewhere the interest is in its development

An interest information unit (IIU) is one distinct fact stated about an interest beyond its name. The coding scheme has 28 codes in four categories.

Study 1: asking

Existing transcripts of college students tutored by an interest-based LLM tutor (Dr. Abhishek Kulkarni’s dataset). Every participant named an interest.

college students tutored by an interest-based LLM tutor

35

units, on average, in the reply to the tutor’s opening question

2.5

said nothing beyond the interest’s name

29%

participants’ disclosures placed the interest in an individual (later) phase

12 of 35

How much of what students said was knowledge of the interest’s domain, before and after the tutor used the interest in teaching:

• Reply to the tutor’s opening question · 86 units

  • 9 domain knowledge (10%)
  • other information

• Added later, while using the interest in teaching · 33 units from 14 participants

  • 29 domain knowledge (88%)
  • other information

• More disclosure went with higher perceived usefulness of the interest-based session (r = .34; 0.11 points per unit on a seven-point scale, adjusted), but not significantly with motivation, relatedness, or quiz scores. Evidence of a later, individual phase also correlated with perceived usefulness.
• Two coders double-coded 25 transcripts and agreed on 90.7% of 700 code-by-transcript decisions; code-level Cohen’s κ was .51, and 65 case-level disagreements were resolved under written rules.

Study 2: reading records (IEEE ICALT 2026)

A within-subjects study. An LLM extracted interests, with explanations, from each student’s samples of browsing history and social-media posts, and students confirmed or rejected each one.

Interests students confirmed (precision)

• Browsing history69.8%
• Social-media posts56.3%

Overlap with what students reported themselves (Jaccard)

• Browsing history.44
• Social-media posts.24

Grounded units in the LLM’s explanations, per confirmed interest

• Browsing history1.58
• Social-media posts2.14

p = .023

Each panel has its own scale, starting at zero.

participants, before and after data-quality screening

77 → 32

overlap between the two sources (Jaccard)

.31

of each student’s validated interests, on average: confirmed interests the two sources added that were missing from the self-report form

38.2%

Posts also placed a smaller share of interests in the early, situational phases, and 54% of the interests missing from self-report were in the two situational phases. Units count as grounded under the grounding check. Privacy protocol: participant self-curation, sensitive-content filtering, and non-retention. The study was supported by NSF CAREER #1942937.

What it adds up to

“Browsing history suits finding the right interest, posts suit learning about it, and both suit covering more interests.” “Asking gives the interest in the student’s own words, teaching draws out what the student knows about it, and data sources, used with consent, add interests the student did not name.”

The appendices publish the materials for reuse: the coding scheme with worked examples and rules, the survey instrument, and the extraction prompts with synthetic examples.

Dissertation: Toward Scalable Interest-Based Learning: Comparing Data Sources for AI-Based Automated Identification of Students’ Personal Interests. Ph.D. in Human-Centered Computing, University of Florida. Committee: Sharon Lynn Chu (chair), Juan E. Gilbert, Lisa Anthony, Pavlo Antonenko.