Search Results

ABSTRACT

Digital tutoring tools are increasingly used in schools to support mathematics learning, but empirical evidence remains limited on how diagnostic and instructional functions can be coordinated to provide personalised support. This study examined a dual-agent intervention in which a diagnostic agent generated provisional learning-profile classifications (diverging, assimilating, converging, and accommodating) and a personalised tutoring agent delivered profile-responsive, stage-specific guidance within Polya’s (1945) problem-solving framework; a general tutoring agent provided non-profile-responsive guidance in the control condition. The quasi-experimental study involved 206 Chinese grade 5 students, including 99 in the experimental group and 107 in the control group. The pre-test and post-test used two parallel mathematics tasks evaluated using the same analytic rubric, and learning gain was calculated from these matched assessment components. The experimental group achieved a significantly larger mean gain than the control group (5.14 vs. 0.94, t(204) = 4.36, p < .001, Hedges’ g = .61). A baseline-adjusted sensitivity analysis confirmed the overall group effect (p < .001). After Holm correction for four profile-specific comparisons, only the assimilating-profile contrast remained significant (adjusted p = .0496). However, the omnibus group-by-profile interaction was not significant (p = .931), warranting caution in interpreting profile-specific differences. Overall, the findings support separating diagnostic and tutoring roles and structuring artificial intelligence (AI) guidance around mathematical problem-solving stages.

Keywords: learning profiles, pedagogical AI agents, personalised mathematical tutoring, Polya’s problem-solving model, quasi-experimental design