IMT Atlantique's didactical-driven teacher assistant achieves 73% precision on 195 questions

Educational chatbots powered by large language models often delegate pedagogical decisions implicitly to the LLM, making tutoring strategies hard to trace, evaluate, and reproduce. Researchers from IMT Atlantique address this by presenting a didactical-driven teacher assistant for a French-language dimensional modeling course that operates without commercial LLM budget or GPU infrastructure. The architecture formalizes instructor pedagogical reasoning into deterministic modules that handle intent detection, concept linking, and didactic approach selection before any text is generated; the LLM acts solely as a linguistic executor. This approach ensures full traceability of tutoring decisions.

Evaluation on 195 authentic student questions revealed two key findings. First, standard semantic retrieval alone does not reliably recover the pedagogically required content, justifying the upstream orchestration strategy. Second, compared to free-tier LLMs whose detection performance varies widely and produces errors silently, the deterministic pipeline achieves high pair precision (73%) with full traceability and explicit abstention. However, its limited coverage confirms that the detection strategy requires further refinement. The work, presented at the CSEDU 2026 conference, demonstrates a practical path to building LLM-based educational tools that are both effective and accountable.