LLM teaching effectiveness varies substantially across subject domains, strongest where knowledge points apply directly and weakest in physics and chemistry¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q2
Currency · newest study 2026 · checked against its sources 2026-10-19 (automated, 2 of 2 entries) · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Across six additional subjects, teaching outcomes rank Mathematics, History, and Politics strongest, Geography and Biology intermediate, and Physics and Chemistry weakest. → Zheng Li 2026 q2 i? In physics and chemistry, teaching often fails to help or even hurts: e.g., Qwen3-235B shows -6.30 accuracy change in Chemistry and -5.65 in Physics versus gains in Politics and History. → Zheng Li 2026 (2)

Evidence¶

Zheng Li 2026¶

Zheng Li, Siyao Song, Jingyuan Ma, Rui Li, Ying Zeng, Minghao Li, Zhifang Sui. (2026). TeachBench: A Syllabus-Grounded Framework for Evaluating Teaching Ability in Large Language Models. https://arxiv.org/abs/2601.21375 doi:10.48550/arxiv.2601.21375

q2 · i? · causal · r2

Cross-domain experiment (Table 4) evaluating teacher agents on Biology, Chemistry, Physics, Politics, History, and Geography subsets, showing "substantial variation in teaching outcomes across domains" and domain dependence of LLM teaching capability.

"Across domains, we observe a clear performance ranking. Mathematics, History, and Politics exhibit the strongest teaching outcomes, followed by Geography and Biology, while Physics and Chemistry show the weakest performance."

Zheng Li 2026 (2)¶

Zheng Li, Siyao Song, Jingyuan Ma, Rui Li, Ying Zeng, Minghao Li, Zhifang Sui. (2026). TeachBench: A Syllabus-Grounded Framework for Evaluating Teaching Ability in Large Language Models. https://arxiv.org/abs/2601.21375

q2 · i? · design · r2

Table 4 row for Qwen3-235B: accuracy changes of -6.30 in Chemistry and -5.65 in Physics versus +2.95 in Politics and +4.65 in History, illustrating that teaching can reduce student accuracy in integration-heavy domains. Values are descriptive; no significance tests are reported.

"Qwen3-235B 77.12 -1.22 65.85 -6.30 66.28 -5.65 81.69 2.95 82.03 4.65 77.14 -2.59"

Discussion¶

Related Claims¶

• 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
• Harm profiles are subject-dependent: metacognitive harm crosses 90% in all three subjects in multi-turn, peaking at 97.44% for Qwen2.5-32B in Mathematics — related
• VietEduQwen improves over its base model across all eight exam subjects, with the largest gains in STEM disciplines, and its residual errors concentrate in multi-step computation and cross-concept integration — related
• LLMs struggle most with Chemistry riddles, with best EM accuracy at or below 75% versus above 90% in other subjects — related
• Multi-turn teaching by LLM teacher agents yields larger student gains than knowledge-point exposure alone — related
• DeepTutor's interactive tutoring gains are stable across five university-level domains, varying only 0.16 points in overall quality — related
• Positive subject-specific effects of Medly in Physics, Chemistry and Biology, with no evidence the effect differed across subjects — related
• SFT on K12-Train improves out-of-domain subjects (Chinese, Humanities Math), suggesting acquisition of a transferable structurally-grounded answer style — related
• Concept-mapping effect sizes vary by science subject (physics/earth science g = 1.040, biology g = 0.671, chemistry g = 0.590) but not significantly — related