Evidence-Aware Agentic Generation of Educational Content: Coverage, Grounding, Repair and Computational Trade-offs

This study investigates the trade-offs between teaching-point coverage, evidence grounding, verification-driven repair and computational efficiency in AI-generated educational content. It examines whether requirement planning, evidence auditing and claim verification improve educational completeness and support while preserving previously addressed teaching points. Six architectures are compared: direct generation (B1), web-only retrieval-augmented generation (B2), multi-source RAG (B3), agentic retrieval (B4), verification-enhanced agentic generation (B5) and a proposed requirement-aware agentic workflow (P1). The experiment evaluates ten educational questions across different subject domains, producing 60 final outputs with one generation per architecture–question pair. All systems share a generation model; retrieval-based workflows use frozen question-specific evidence pools. A separate evaluator assesses saved outputs using a common teaching-point checklist. Evaluation covers question coverage, reference evidence support, narration quality, Grounded Requirement Coverage, Context Precision@5, Context Recall, citation quality, Repair Effectiveness, Requirement Preservation Rate, generation cost and latency. Retrieval-based systems achieve mean reference evidence support of 90.7–94.1%, compared with 56.5% for direct generation. P1 achieves the highest mean question coverage among retrieval-based systems (86.9%) and the highest mean context recall (83.3%). B4 achieves the highest observed mean Grounded Requirement Coverage (67.2%), followed by P1 (66.7%); however, no pairwise difference in this metric remains statistically significant after the reported multiple-comparison correction. Context recall is more strongly associated with grounded requirement coverage than top-five passage relevance under the study’s automated evaluation protocol. Among cases where repair changes the output, mean Repair Effectiveness is −0.46 percentage points for B5 and −1.00 percentage points for P1. This metric measures changes in coverage and narration, rather than changes in factual support. Mean Requirement Preservation Rate is 76.7% for B5 and 75.9% for P1, indicating that revisions do not consistently preserve the previous coverage scores of addressed teaching points. P1 incurs the highest generation cost and latency, averaging USD 0.15902 and 224.8 seconds per output, compared with USD 0.01448 and 23.8 seconds for B4. Under explicitly stated qualitative priorities emphasising grounded requirement coverage, evidence support, completeness and computational efficiency, the paper recommends B4 as a practical starting point for resource-conscious educational content generation. This recommendation is an exploratory, conditional judgement, not a statistically established overall ranking. The study contributes an implemented requirement-aware workflow, a joint operational evaluation of coverage and grounding with repair preservation, and an auditable comparison of six architectures. Its findings show that increased agentic complexity does not consistently improve measured outcomes and support evaluating completeness, evidence support, revision preservation and computational resources together. The experiment relies on automated assessment, a limited benchmark and a single generation per architecture–question pair. Differences in workflow components, evidence budgets and synthesis representations prevent causal attribution to individual design choices. The findings do not establish independently validated factual accuracy, classroom effectiveness or improvements in student learning outcomes. Keywords: Agentic AI; Retrieval-Augmented Generation; Artificial Intelligence in Education; Educational Content Generation; Large Language Models; Evidence Grounding; Requirement Coverage; Grounded Requirement Coverage; Claim Verification; Verification-Driven Repair; Requirement Preservation; Automated Evaluation; Computational Trade-offs.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24245960
Primary Topic
Artificial Intelligence in Education
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Evidence-Aware Agentic Generation of Educational Content: Coverage, Grounding, Repair and Computational Trade-offs

RAJU BADDELA
Zenodo (CERN European Organization for Nuclear Research)
Artificial Intelligence in Education
preprint

Evidence-Aware Agentic Generation of Educational Content: Coverage, Grounding, Repair and Computational Trade-offs

RAJU BADDELA
preprint en

Abstract

This study investigates the trade-offs between teaching-point coverage, evidence grounding, verification-driven repair and computational efficiency in AI-generated educational content. It examines whether requirement planning, evidence auditing and claim verification improve educational completeness and support while preserving previously addressed teaching points. Six architectures are compared: direct generation (B1), web-only retrieval-augmented generation (B2), multi-source RAG (B3), agentic retrieval (B4), verification-enhanced agentic generation (B5) and a proposed requirement-aware agentic workflow (P1). The experiment evaluates ten educational questions across different subject domains, producing 60 final outputs with one generation per architecture–question pair. All systems share a generation model; retrieval-based workflows use frozen question-specific evidence pools. A separate evaluator assesses saved outputs using a common teaching-point checklist. Evaluation covers question coverage, reference evidence support, narration quality, Grounded Requirement Coverage, Context Precision@5, Context Recall, citation quality, Repair Effectiveness, Requirement Preservation Rate, generation cost and latency. Retrieval-based systems achieve mean reference evidence support of 90.7–94.1%, compared with 56.5% for direct generation. P1 achieves the highest mean question coverage among retrieval-based systems (86.9%) and the highest mean context recall (83.3%). B4 achieves the highest observed mean Grounded Requirement Coverage (67.2%), followed by P1 (66.7%); however, no pairwise difference in this metric remains statistically significant after the reported multiple-comparison correction. Context recall is more strongly associated with grounded requirement coverage than top-five passage relevance under the study’s automated evaluation protocol. Among cases where repair changes the output, mean Repair Effectiveness is −0.46 percentage points for B5 and −1.00 percentage points for P1. This metric measures changes in coverage and narration, rather than changes in factual support. Mean Requirement Preservation Rate is 76.7% for B5 and 75.9% for P1, indicating that revisions do not consistently preserve the previous coverage scores of addressed teaching points. P1 incurs the highest generation cost and latency, averaging USD 0.15902 and 224.8 seconds per output, compared with USD 0.01448 and 23.8 seconds for B4. Under explicitly stated qualitative priorities emphasising grounded requirement coverage, evidence support, completeness and computational efficiency, the paper recommends B4 as a practical starting point for resource-conscious educational content generation. This recommendation is an exploratory, conditional judgement, not a statistically established overall ranking. The study contributes an implemented requirement-aware workflow, a joint operational evaluation of coverage and grounding with repair preservation, and an auditable comparison of six architectures. Its findings show that increased agentic complexity does not consistently improve measured outcomes and support evaluating completeness, evidence support, revision preservation and computational resources together. The experiment relies on automated assessment, a limited benchmark and a single generation per architecture–question pair. Differences in workflow components, evidence budgets and synthesis representations prevent causal attribution to individual design choices. The findings do not establish independently validated factual accuracy, classroom effectiveness or improvements in student learning outcomes. Keywords: Agentic AI; Retrieval-Augmented Generation; Artificial Intelligence in Education; Educational Content Generation; Large Language Models; Evidence Grounding; Requirement Coverage; Grounded Requirement Coverage; Claim Verification; Verification-Driven Repair; Requirement Preservation; Automated Evaluation; Computational Trade-offs.

Zenodo (CERN European Organization for Nuclear Research)
Artificial Intelligence in Education
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon