AREX: A Recursively Self-Improving Deep Research Agent from BAAI

Loading video

Loading video

AREX alternates an inner research loop that builds a provisional answer with an outer self-improvement loop that audits every claim constraint-wise and turns unresolved ones into the next research objective. A learned context-update tool compresses history into a compact improvement state. Ships as a dense 4B model and a 122B-A10B MoE.