GAIMHE aims to design and evaluate hybrid EdTech architectures that combine:
- Frugal and pedagogically robust Intelligent Tutoring Systems (ITS) for macro-level orchestration
- Generative AI (LLM/SLM) for micro-level personalized feedback and content generation
This hybrid approach leverages the strengths of both paradigms to scale learning experiences, ensure pedagogical quality, and reduce computational costs compared to full-LLM solutions.
- Generation of large banks of pedagogical exercises, feedback, and hints using LLMs (e.g., Llama 3, GPT-4, Claude, Gemini)
- Orchestration with classical AI techniques
- Human expert validation of generated content
- Content stored and reused to minimize live calls to LLMs
- Targeted use of SLM such as Mistral 7B, Llama 3 8B, or Phi 3 mini
- Activation in specific situations:
- When a learner is stuck after multiple attempts
- For open-ended or complex activities (e.g., text production, metacognition tasks)
- Ensures personalized and adaptive support while preserving efficiency
- Pre-generation of pedagogical datasets using LLMs and expert validation
- Development of annotation protocols and codebooks to ensure quality and reliability
- Automatic generation and annotation of large-scale synthetic datasets using prompt engineering, RAG, in-context learning, and RLHF
- Collaboration with OpenLLM-France for SLM fine-tuning and benchmarking
The hybrid approach is estimated to be around 30 times more efficient than full-LLM approaches.
It is scalable to millions of learners with significantly reduced environmental impact.
- EvidenceB
- Inria
- ClassCode / OpenLLM-France
- Région Île-de-France
- Le café pédagogique
- Scaleway