Multi-agent AI debate system generates scientific hypotheses from 500 papers

Modern scientific discovery suffers from fragmented knowledge, especially in battery materials where performance, interfaces, and manufacturing must be optimized simultaneously. To address this, researchers from multiple institutions developed the Multi-Persona Debate System (MPDS), a framework that assembles up to 500 papers into a literature snapshot, grounds multiple AI agents in role-specific evidence pools (e.g., electrochemist, manufacturing engineer), and runs a three-round structured debate. A moderator agent then synthesizes the discussion while preserving citation traceability. The system was evaluated under a temporally controlled protocol that excluded direct access to target papers, using two held-out battery case studies (sodium-ion anodes and all-solid-state cathodes) and a blinded comparison across 30 matched cases.

MPDS outperformed simpler baselines by generating more mechanistically explicit and process-aware proposals. The team introduced Integrative Hypothesis Quality scoring; in ablation studies, MPDS achieved the highest mean score across five conditions, with its largest advantage in cross-perspective integration. A laboratory follow-up confirmed its utility as a diagnostic aid for identifying practical workflow bottlenecks. The results demonstrate that structured multi-agent debate over large literature snapshots improves hypothesis formation under coupled engineering constraints, offering a reusable workflow for text-intensive scientific discovery in materials science and beyond.