Instruction-tuning a student model on a single misconception overgeneralizes the error across problems and degrades correct-solving accuracy¶

Claim · All claims
Evidence · 1 study · causal r2 · q2
Currency · newest study 2026 · checked against its sources 2026-10-20 (automated, 1 of 1 entry) · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Naive single-misconception training causes LLMs to overgeneralize erroneous reasoning to problem types where the misconception should not apply, degrading correct solving. → 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

Instruction-tuning experiments across Llama-3.1-8B-Instruct, Phi-4-4B-Mini, and Qwen-3-4B-Instruct on single misconceptions showed that as Misconception Accuracy rose, Overall Correct Accuracy and CANA declined; the article reports "a dramatic decline in OCA andCANA" with no effect size printed.

"Instead, naive training causes models to overgeneralize erroneous reasoning patterns, degrading their ability to solve even problems where the misconception should not apply."

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
• LLM error simulation rarely fails, but the model is less likely to reproduce a targeted distractor on problems with longer correct solutions — related
• Learner errors arise from interlingual transfer, intralingual overgeneralization, cultural interference, and communicative strategies — related
• On the misconception-annotated math dataset, LLMs predominantly follow a solve-first, error-injection process aligned with misconception-based distractor design — related
• Incorporating example problems does not necessarily improve LLM teaching and can degrade it via a shift to example-specific error correction — related
• Self-explanation promotes transfer of general patterns but can reduce recall of details and cause overgeneralization — related
• Tutor models trained on ten misconceptions simultaneously learn each misconception while overall correct accuracy remains stable or improves (93% to 98%) — related