From Scoring to Explanations: Evaluating SHAP and LLM Rationales for Rubric-based Teaching Quality Assessment
Ivo Bueno, Babette Bühler, Philipp Stark, Tim Fütterer, Ulrich Trautwein, Dorottya Demszky, Heather Hill, Enkelejda Kasneci · arXiv (2026) · doi:10.48550/arxiv.2606.05180 · gemini-3.6-flash
TL;DR
This study tests whether artificial intelligence models can explain why they give specific quality ratings to classroom teaching transcripts. While smaller fine-tuned AI models gave more accurate score predictions, SHAP statistical explanations proved far more reliable than the explanations written by large language models. The findings indicate that explanations generated directly by large language models should not be trusted as true justifications for automated educational scores.
Method and sample: The study analyzed 6,005 15-minute classroom transcript segments from 1,600 elementary mathematics lessons in the NCTE corpus (4,775 train, 1,230 test), annotated on the 1–7 CLASS Quality of Feedback dimension. Fine-tuned encoder models (BERT, ALBERT, RoBERTa, DeBERTaV3 base/large) were compared against 4-bit quantized open-source instruction-tuned LLMs (Llama 3.1 8B/70B, Mixtral 8x7B/8x22B, Qwen 3 4B/30B/235B, Mistral Small/24B). Baselines included constant mean/median score baselines and length-matched random sentence deletion.
#Automated Scoring#Rubric Assessment#LLM Rationales#SHAP#Explanation Faithfulness#Classroom Assessment Scoring System#Teaching Quality Assessment