Answers are grounded in this paper and cite their sources.
Collaborative learning scripts are usually fixed, and collaboration dashboards do not necessarily prompt teachers to intervene. We present a hybrid intelligent decision-support architecture that diagnoses five group dysfunctions, explains each alert and proposes one script adaptation from a closed catalogue, with an open simulation benchmark providing planted dysfunctions and a per-group oracle. It combines fuzzy inference, a lexical content classifier, calibrated gradient-boosting detectors with SHAP explanations, expected-benefit operator selection and NSGA-II regrouping. All evidence comes from controlled simulations; no teacher used the system, and learning outcomes were not evaluated. Over 20 development datasets, the detectors reached a macro-F1 of 0.591, and the fuzzy indicators were practically equivalent to their absence in accuracy. Because development data, including test windows of one seed, were inspected while the system was built, these results are not an independent confirmation; a pre-specified evaluation of the frozen system on 20 new datasets replicated 14 of 15 endpoints (macro-F1 0.590), but the defect rate of filtered explanations doubled (0.029 to 0.059). Performance did not survive distribution shift: macro-F1 fell from 0.586 to 0.297 under a changed generating mechanism, the full model fell below the interaction-only model on a held-out topic, and misconception F1 on unseen paraphrases (0.146) was no better than flagging every window. A training-selected operating point achieved 0.399 of the oracle’s simulated harm reduction with 18.7% unneeded interventions, and detector-triggered regrouping increased simulated harm. The benchmark serves as a pre-deployment stress test; the scores are upper bounds, not evidence of teaching benefit.
No takes yet. Share an insight, caveat, or question.
Stamou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: