Distilling a multi-order hierarchical teacher into a smaller zeroth-order student lets the student outperform the teacher, and improves 4.1 percentage points in accuracy over STGCN++¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? In ablations, distilling the multi-order hierarchical teacher into a single zeroth-order student allows the student to outperform the teacher, and imposing the smaller-student condition further enables it to outperform the much larger hierarchical teacher. → Paritosh Parmar 2026 q2 i? Compared with the base STGCN++ model, the proposed method improves accuracy by 4.1 percentage points and macro-F1 by 5.2 percentage points, mainly via better recognition of kick, sit, and run. → Paritosh Parmar 2026 (2)

Evidence¶

Paritosh Parmar 2026¶

Paritosh Parmar, Landy Lan, Hong Yang, Chen Yi, Chiat Pin Tay. (2026). Robust and Efficient Motion Reasoning for Privacy-Aware Classroom Incident Recognition. arXiv preprint. https://arxiv.org/abs/2608.05115 doi:10.48550/arxiv.2608.05115

q2 · i? · design · r2

Ablation study (Figure 8, panel 3) comparing zeroth-order teacher, multi-order hierarchical teacher, and distilled students with and without the size condition. The authors hypothesize the student internalizes multi-order motion reasoning; accuracies only, no effect size.

"we observe that distilling a multi-order hierarchical teacher into a single zeroth-order student allows the student to outperform the teacher. Third, imposing the condition in Equation 11 further enables the student to outperform the much larger hierarchical teacher."

Paritosh Parmar 2026 (2)¶

Paritosh Parmar, Landy Lan, Hong Yang, Chen Yi, Chiat Pin Tay. (2026). Robust and Efficient Motion Reasoning for Privacy-Aware Classroom Incident Recognition. arXiv preprint. https://arxiv.org/abs/2608.05115

q2 · i? · design · r2

Ablation-reported gains over the STGCN++ base model: "improves accuracy by 4.1 percentage points and macro-F1 by 5.2 percentage points", attributed in the Appendix mainly to better kick, sit, and run recognition and reduced confusion of kick and sit with the other class.

"Furthermore, compared with the base model, STGCN++ [10], the proposed method improves accuracy by 4.1 percentage points and macro-F1 by 5.2 percentage points."

Discussion¶

Related Claims¶

• A distilled multi-order kinematic model outperforms pose-based baselines on a synthetic classroom incident dataset while using one-tenth as many parameters as PoseC3D — related
• Models trained on synthetic classroom incident data transfer zero-shot to real-world footage, with the proposed method best at 63.41% though a synthetic-to-real domain gap remains
• Classroom incident recognition is substantially harder than established general action-recognition benchmarks, with PoseC3D dropping from over 94% accuracy to 70.54% — related