Advancing Bloom’s Cognitive Levels through Socratic Dialogue: A DPO-Based Multi-Agent AI Framework for K-12 Learners

The widespread adoption of Generative AI technologies in K-12 education is accompanied by tremendous opportunities for reform as well as significant challenges, including the tendency of students to look for quick answers with less intellectual effort. To address this problem, we suggest a specifically designed Socratic multi-agent conversational AI system, designed and tuned with care with Direct Preference Optimization (DPO) to support higher-order thinking processes as defined in Bloom's Taxonomy. Our strategy was to create a carefully curated dataset of 734 pairs of dialogues, based on the NCERT Science Class 7 textbook "Curiosity", wherein brief factual answers were studiously avoided in favour of longer Socratic dialogue generated using the Llama 3 large language model (LLM). We used the Qwen 2.5 language model (with 7 billion parameters) to perform parameter-efficient fine-tuning with Low-Rank Adaptation (LoRA), directly optimizing the model with human-annotated preference data comparing Socratic dialogues with direct factual answers. The developed Socratic multi-agent system, built through the SmolAgents framework offered by Hugging Face, consists of a master agent to coordinate the interactions of a Socratic questioning agent and a motivational support agent built to encourage student motivation. Experiments demonstrated a significant improvement in the analytical thinking, creativity, and self-directed problem-solving skills of students, reaffirming the potential of this new AI paradigm to foster responsible and meaningful educational debate. The model, dataset, and accompanying code have been released publicly to facilitate reproducibility and further research.