Without step-level supervision, neither the student model nor the tutor model can acquire misconceptions regardless of training data size¶

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? Training on final incorrect answers alone leaves MA below 30% across all three base models even with up to 1600 (student) or 800 (tutor) samples. → 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

No-steps ablation repeating the acquisition experiments with final answers only; several misconceptions (M3, M5, M8) achieved less than 1% accuracy, and for the tutor model without steps OCA also degraded as training progressed.

"Across all three base models and all misconceptions, MA remained below 30% even with substantial training data (up to 1600 samples for the student model, 800 for the tutor model)."

Discussion¶

Related Claims¶

• On the PRM800K validation dataset, detection rates dropped by 10.5 percentage points when final answers were correct, consistent with the correct answer trap extending beyond the Eedi data — related
• Classroom-scale data (5-80 samples per misconception) is insufficient for reliable tutor-model misconception acquisition — related
• Tutor models trained on ten misconceptions simultaneously learn each misconception while overall correct accuracy remains stable or improves (93% to 98%) — related