Model scale does not consistently improve pedagogical safety: within the Qwen2.5 family the 72B model beats the 7B on only 17 of 33 subject-dimension pairs¶

Claim · All claims
Evidence · 1 study · 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? Scaling from 7B to 72B within the Qwen2.5 family improves safety on some dimensions but regresses on others, with the 72B model lower than the 7B on only 17 of 33 single-turn subject-dimension pairs. → Rima Hazra 2026

Evidence¶

Rima Hazra 2026¶

Rima Hazra, Bikram Ghuku, Ilona Marchenko, Yaroslava Tokarieva, Sayan Layek, Somnath Banerjee, Julia Stoyanovich, Mykola Pechenizkiy. (2026). SAFETUTORS: Benchmarking Pedagogical Safety in AI Tutoring Systems. arXiv preprint. https://arxiv.org/abs/2603.17373 doi:10.48550/arxiv.2603.17373

q2 · i? · design · r2

Comparison of Qwen2.5 family models (7B, 14B, 32B, 72B), which share architecture and training recipe, on single-turn SAFETUTORS harm rates. The article reports the 72B "achieves a lower harm rate than the 7B on only 17 pairs, a higher rate on 14, and ties on 2"; no effect size is printed.

"Across all 33 subject–dimension pairs in single-turn, the 72B model achieves a lower harm rate than the 7B on only 17 pairs, a higher rate on 14, and ties on 2."

Discussion¶

Related Claims¶

• CDPK performance scales with model size, with a sharp Pareto-frontier drop-off below around 8B parameters — related
• Model family and instruction-tuning approach appear better predictors of tutoring quality than parameter count alone — a broader claim this one bears on
• Single-turn tutors disclose answers pervasively while challenge scores are near zero across every model and subject — related
• Education-oriented instruction tuning consistently improves curriculum grounding, K-12 problem solving, and pedagogical tutoring at every evaluated model scale — related
• Model scale alone does not induce Socratic behavior: unaligned Qwen-72B achieves 0% scaffolding effectiveness and 96.7% keyword leakage — a narrower finding that bears on this claim
• Emotion features improve tutor performance only when grounded in education-theory-based pedagogical strategies; emotion information alone misguides a smaller model — related
• Multi-turn interaction amplifies pedagogical harm: mean harm rises 6-11 percentage points across subjects and pedagogical-relationship harm surges from 17.7% to 77.8% — related
• No evaluated LLM tutor is reliably safe: every model exceeds 60% harm rate on at least five risk categories in single-turn and six in multi-turn evaluation — related