Learned tutoring policies show different classroom learning and override patterns
Source check: Source checked. Not independent scientific confirmation.
Finding: In a probability-tutor classroom study, an apprenticeship-learned policy overrode student choices less often than a conservative-Q-learning policy (28.2% versus 66.7% of recorded decisions). Among non-ceiling students, the apprenticeship group had a higher mean standard posttest score (.908 versus .842), with a marginal planned one-sided pretest-adjusted contrast. Small samples, repeated decisions and an augmented historical control limit broader superiority claims.
Technology: StuEDM; energy-based distribution matching; conservative Q-learning
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- An energy-based neural policy learned from student demonstration occupancy selects presentation modes and overrides conflicting choices during new classroom use; measured learning and choice patterns are the represented field result. Formatting assistance by a separate language model is unrelated to this contribution.
- Test and result
- StuEDM was trained on 244 earlier student trajectories; StuCQL used 2,716. The Spring 2026 live study had 104 completers. Removing perfect-pretest students left 26 StuEDM and 27 StuCQL students; the random-mode Expert comparator was augmented with 11 prior-semester students. Table 7.1 reports a planned one-sided adjusted contrast p=.048 for StuEDM versus StuCQL; the three-group omnibus test was not significant.
- Limitation / still unknown
- The Expert arm is partly historical, and non-significant baseline tests do not prove population equivalence. Other direct learning comparisons against StuCQL were not significant. Multiple directional outcomes, ceiling exclusions, a single course and small samples limit inference. Override counts include repeated decisions within students; decision-level significance is not independent-person replication. Longer tutor time is measured engagement, not proof of benefit. Earlier dissertation chapters reproduce published methods, which are not represented as new here; October 20 is the located arXiv version date, not a claim of absolute first circulation.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/html/2610.28716v1 · https://arxiv.org/abs/2610.28716v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: Original v1 dissertation front, authorship/version distinctions and Chapter 7 learned-policy Methods, classroom design, Results and limitations.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 1/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.