Multi-agent debate framework boosts language learning with 90.91% accuracy

Researchers Nicy Scaria, Silvester John Joseph Kennedy, and Deepak Subramani introduced Learning in Blocks, a personalized adaptive learning framework that uses a multi-agent debate system called HeteroMAD (Heterogeneous Multi-Agent Debate) to evaluate language learning progress. Unlike traditional digital curricula that rely on discrete-item quizzes, this framework grounds progression in demonstrated conversational competence using CEFR-aligned rubrics. HeteroMAD operates in two stages: a scoring stage where role-specialized AI agents independently evaluate Grammar, Vocabulary, and Interactive Communication, then debate to resolve conflicting judgments, and a recommendation stage that identifies specific grammar skills and vocabulary topics for targeted review. The system requires learners to demonstrate 70% mastery before advancing, with spaced review targeting identified weaknesses to counter skill decay.

Benchmarked on CEFR A2 conversations annotated by ESL experts, HeteroMAD achieved superior score agreement with a 0.23 degree of variation and recommendation acceptability of 90.91%. An 8-week study with 180 CEFR A2 learners demonstrated that combining rubric-aligned scoring and recommendation with spaced review and mastery-based progression produces better learning outcomes than feedback alone. This approach addresses the common problem of learners advancing despite persistent gaps in using grammar and vocabulary during interaction, offering a more reliable and validated method for scoring open-ended conversations and driving personalized learning paths.