Multi-turn teaching by LLM teacher agents yields larger student gains than knowledge-point exposure alone¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? On the Gaokao Mathematics subset, explicit multi-turn teaching interactions improve student Pass@1 accuracy substantially more than providing knowledge points without interaction (e.g., 7.63 points for the strongest teacher vs. 4.37 for knowledge-only). → Zheng Li 2026 q2 i? Qwen3-235B-A22B-Instruct-2507 achieves the strongest teaching performance among evaluated models, with the largest Pass@1 improvement of 7.63 points. → 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

Simulation-style controlled experiment on the Gaokao Mathematics subset with Qwen2.5-7B-Instruct as student agent. Knowledge-point exposure alone yields a "4.37 points" Pass@1 gain over the no-teaching baseline, which the authors call "only modest benefits".

"Qwen2.5-7B-Instruct(Knowledge) improves Pass@1 accuracy by 4.37 points over the no-teaching baseline. This result indicates that access to structured knowledge alone provides only modest benefits."

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

Same Mathematics-subset evaluation across multiple teacher agents, each running three independent teaching sessions with 64 repeated student evaluations. Qwen3-235B-A22B-Instruct-2507 achieves the "largest Pass@1 im- provement of 7.63 points", and gains from teaching consistently exceed knowledge-only exposure.

"Among the evaluated models, Qwen3-235B- A22B-Instruct-2507 achieves the strongest teach- ing performance, yielding the largest Pass@1 im- provement of 7.63 points."

Discussion¶

Related Claims¶

• Incorporating example problems does not necessarily improve LLM teaching and can degrade it via a shift to example-specific error correction — related
• Teaching primarily helps the student agent realize existing potential rather than raise its absolute performance ceiling — related
• LLM teaching effectiveness varies substantially across subject domains, strongest where knowledge points apply directly and weakest in physics and chemistry — related
• In math problem solving, a tutor-plus-peers LLM arrangement produced the strongest unassisted post-task test accuracy, with accuracy increasing monotonically from Control to Tutor + Peers — related
• Learning By Teaching Improves Tutor Learning — 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
• Example-based instructional activities yield larger effects when delivered jointly by ICT and teachers than by teachers alone — related