The framework is robust to 15% ASR transcription noise and retains 99.6% of its ICS under zero-shot cross-domain transfer to EdNet-Math¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q2
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¶

q2 i? Under 15% injected ASR transcription errors, ICS drops by only 4.1% relative for the proposed model versus drops exceeding 12% for all baselines. → Li 2025 q2 i? Without fine-tuning on EdNet-Math, the model retains 99.6% of its TeacherEval-2023 ICS score, outperforming baselines by 6–11% on ICS and ExpScore. → Li 2025 (2)

Evidence¶

Li 2025¶

Li, Y.; Yang, H.; Fang, Q. (2025). Towards Trustworthy and Explainable-by-Design Large Language Models for Automated Teacher Assessment. Information 2025, 16, 882. https://doi.org/10.3390/info16100882

q2 · i? · design · r2

Robustness evaluation (§4.5) under noisy input: 15% ASR transcription errors injected, with the model's ICS degradation described as minimal and visualized in Figure 4, versus baseline drops exceeding 12%.

"When injecting 15% ASR transcription errors, the Inter-Rater Consistency Score (ICS) drops by only 4.1% relative for our model, compared to drops exceeding 12% for all baselines."

Li 2025 (2)¶

Li, Y.; Yang, H.; Fang, Q. (2025). Towards Trustworthy and Explainable-by-Design Large Language Models for Automated Teacher Assessment. Information 2025, 16, 882. https://doi.org/10.3390/info16100882

q2 · i? · design · r2

Cross-dataset transfer test (§4.5, Table 5) with paired t-tests reported significant (p < 0.05) against fine-tuned BERT, Instructor-LM, and GPT-4 zero-shot on ICS and ExpScore.

"For cross-dataset transfer, we evaluate performance without fine-tuning on EdNet-Math. As shown in Table 5, the model retains 99.6% of its TeacherEval-2023 ICS score, outperforming baselines by 6–11% on both ICS and ExpScore."

Discussion¶

Related Claims¶

• Models trained on synthetic classroom incident data transfer zero-shot to real-world footage, with the proposed method best at 63.41% though a synthetic-to-real domain gap remains — related
• A distilled multi-order kinematic model outperforms pose-based baselines on a synthetic classroom incident dataset while using one-tenth as many parameters as PoseC3D — related
• The proposed framework outperforms fine-tuned BERT, Instructor-LM, GPT-4, and RAPT on TeacherEval-2023 across consistency, credibility, calibration, and fairness metrics — related
• Supervised ML/DL Bloom question classifiers trained on one dataset drop substantially in weighted F1 when tested on unseen datasets — related