Tutor models trained on ten misconceptions simultaneously learn each misconception while overall correct accuracy remains stable or improves (93% to 98%)¶

Claim · All claims
Evidence · 1 study · causal 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? With sufficient samples (100-800 per misconception), tutor-model MA increases while OCA remains stable or improves across all three base models. → Naiming Liu 2026

Evidence¶

Naiming Liu 2026¶

Naiming Liu, Xinghe Chen, Richard Baraniuk, Mrinmaya Sachan, and Shashank Sonkar. (2026). Misconception Acquisition Dynamics in Large Language Models. https://arxiv.org/abs/2604.00818 doi:10.48550/arxiv.2604.00818

q2 · i? · causal · r2

Tutor-model instruction-tuning on ten misconceptions (Mset of M1, M2, M6, M8, M11, M12, M19, M20, M21, M22) across Llama-3.1-8B, Phi-4, and Qwen-3; the article reports OCA rising from 93% to 98% with no effect size printed.

"whentrainedwithsufficientsamples(100–800),thetutormodelexhibitsincreasing MA while OCA remains stable or evenimproves, rising from 93% to 98% across all three models (Llama, Phi, Qwen)."

Discussion¶

Related Claims¶

• Mixing correct examples into student-model training resolves the misconception-accuracy trade-off, with correct examples as low as 25% preserving both properties — related
• Classroom-scale data (5-80 samples per misconception) is insufficient for reliable tutor-model misconception acquisition — related
• Without step-level supervision, neither the student model nor the tutor model can acquire misconceptions regardless of training data size — related
• Instruction-tuning a student model on a single misconception overgeneralizes the error across problems and degrades correct-solving accuracy — related