Current LLMs misalign instructional-strategy sensitivity with pedagogical relevance, including hallucinated relevance of irrelevant traits¶

Claim · All claims
Evidence · 1 study · theoretical r2 · q1
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person · recheck by 2028: tool and model accuracy changes quickly

Subclaims¶

q1 i? In leave-one-out analyses, an LLM assigned a meaningless distractor (favorite hobby) a higher impact weight (TVD: 0.273) than nearly all pedagogical traits, and failed to attend to effort regulation (TVD: 0.080). → Naiming Liu 2025

Evidence¶

Naiming Liu 2025¶

Naiming Liu, Brittany Bradford, Johaun Hatchett, Gabriel Diaz, Lorenzo Luzi, Zichao Wang, Debshila Basu Mallick, Richard Baraniuk. (2025). Learning Context: A Unified Framework and Roadmap for Context-Aware AI in Education. https://arxiv.org/abs/2512.24362 doi:10.48550/arxiv.2512.24362

q1 · i? · theoretical · r2

Leave-one-out (LOO) analysis measuring Total Variation Distance between full-context and LOO-context strategy distributions, shown in Figure 3; ground-truth characteristics sourced from the MSLQ, distractors synthetically generated.

"While the model correctly prioritizes impactful traits like the learner's perceived value of the task (TVD: 0.287) and self-efficacy (TVD: 0.207), it fails to attend to other vital factors such as the learner's ability to regulate their effort (TVD: 0.080), treating them as effectively invisible."

Discussion¶

Related Claims¶

• Zero-shot LLM errors follow four recurring patterns: over-interpretation, failure to detect relevant information, hallucination, and failure to generate a response — a broader claim this one bears on
• ChatGPT zero-shot relevance extraction for instruction-quality assessment retrieves mostly irrelevant utterances and overestimates instruction quality — related
• LLM-assisted inductive qualitative coding carries risks of superficial themes, broad or redundant codes, and hallucinated interpretations, so LLMs should augment rather than replace human researchers — related
• Recurring risks of GenAI in STEAM education include hallucination, bias, superficial completion strategies, and compromised assessment validity — a broader claim this one bears on
• General-purpose LLMs show accuracy and reliability limits, including hallucinations and task-specific accuracy below perfect, making them risky as expert advice sources — a broader claim this one bears on
• LLMs in educational applications hallucinate in 15–25% of responses and can amplify bias against non-native language inputs — a broader claim this one bears on