Mixing correct examples into student-model training resolves the misconception-accuracy trade-off, with correct examples as low as 25% preserving both properties¶

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? Adding correct examples during student-model instruction-tuning recovers CANA while keeping Misconception Accuracy high (90-95%). → 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

Mixed-training experiments (nc/nm in {0.25,0.5,1.0}) on Llama-3.1-8B-Instruct, with similar patterns for Phi-4-4B-Mini and Qwen-3-4B-Instruct; for M8, "MA remained consistently high, increasing from 90.2% at 400 samples to 94.5% at 3200 samples."

"MA (blue) remains high ( 90-95%) demon- strating successful misconception acquisition, while CAN A (red) rapidly recovers with just 200-400 correct examples."

Discussion¶

Related Claims¶

• Instruction-tuning a student model on a single misconception overgeneralizes the error across problems and degrades correct-solving accuracy — related
• Tutor models trained on ten misconceptions simultaneously learn each misconception while overall correct accuracy remains stable or improves (93% to 98%) — related