Learning in Blocks: Multi-Agent Framework for Adaptive Language Learning

A new framework called Learning in Blocks uses multi-agent debate to assess conversational language proficiency based on CEFR-aligned rubrics, moving beyond traditional quiz-based progression. The system employs heterogeneous multi-agent debate (HeteroMAD) in two stages: scoring and review. In the scoring stage, role-specialized agents independently evaluate Grammar, Vocabulary, and Interactive Communication, then debate to resolve conflicting judgments before a judge synthesizes consensus scores. This approach aims to ground progression in demonstrated conversational competence rather than discrete-item quiz performance, addressing the limitation of current digital curricula where learners can advance despite persistent gaps in applied grammar and vocabulary use. The framework leverages recent advances in LLM-based judging to score open-ended conversations, requiring reliable and validated scoring protocols. The work is published on arXiv under identifier 2604.22770.