Ablations show the trustworthiness module contributes most to performance, and the explanation generator cuts human review time by 41%¶

Claim · All claims
Evidence · 1 study · 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? Removing the trustworthiness constraints causes the most severe performance drop (ICS −9.2%, ExpScore −14.7%), followed by disabling hierarchical attention (ICS −6.8%); the explanation generator reduces human review time by 41%. → Li 2025

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

Component-wise ablation study (§5.6, Table 11) on the TeacherEval-2023 test set, removing each module individually; review-time impact is +63% for the trustworthiness module, +38% for hierarchical attention, and +112% for the explanation generator.

"Removing the trustworthiness constraints causes the most severe performance drop (ICS −9.2%, ExpScore−14.7%), followed by disabling hierarchical attention (ICS−6.8%). The dynamic explanation generator, while contributing modestly to raw accuracy (+3.1% ICS), reduces human review time by 41% through structured justifications."

Discussion¶

Related Claims¶

• Removing the cognitive load module causes the largest ablation performance drop, while state fusion has a smaller effect — related
• Ablations show each CogEvolution module contributes: removing ICAP perception, structured retrieval, or evolutionary update degrades mistake precision, learning-curve fit, and alignment — related
• In the EduAgent framework, removing past cognitive states from memory significantly reduced simulation performance — related
• Ablation study shows metadata-enriched embeddings produce the largest accuracy drop, while attention masking, KV caching and dynamic batching primarily support inference efficiency — related
• Each pedagogical verification component contributes to a distinct quality dimension: layout verification most strongly — related
• Federated aggregation is the primary driver of performance gains, contributing an average 11.99% improvement — related
• Static Knowledge Grounding and Dynamic Personal Memory are complementary: removing both yields the largest quality degradation — related