Work

A Fault-Tolerant Cross-Lingual RAG Architecture for Reliable Curriculum Graph Generation

Year: 2026
Type: conference-paper
Abstract: Large language models (LLMs) can draft educational programmes quickly, but direct generation is unreliable when the output must satisfy structural constraints such as parseable JSON, acyclic prerequisite relations, semester-order consistency, credit balance, learning-outcome coverage, and recoverability after workflow failures. This paper presents a faulttolerant cross-lingual retrieval-augmented generation (RAG) architecture for reliable curriculum graph generation. The system treats the LLM as a noisy proposal generator rather than a final decision maker. It combines evidence-guided proposal generation with deterministic schema repair, prerequisite-graph validation, semester correction, learning-outcome coverage repair, and persistent task recovery. To evaluate the reliability layer without exposing private institutional data, we build a controlled, seeded curriculum-graph stress-test benchmark with 80 programme intents across eight disciplines, 3,120 courses, 1,197 learning outcomes, 4,765 prerequisite edges, and 3,033 learning-outcome-to-course mappings. Across five random seeds, the proposed pipeline substantially improves structural validity and failure recovery compared with LLM-only and RAG-only variants. It also provides competitive prerequisite-edge recommendation with cycle-free outputs and reduced LLM-call cost. The results do not claim pedagogical optimality; instead, they show that deterministic validation and recoverable execution are necessary control layers for reliable LLM-assisted curriculum graph generation.
Language: English
Subfield: Education
Open Access status: hybrid
OpenAlex ID: W7225993695
FWCI:
Cites:
Cited by:
Related to: 10