AREX: Towards a Recursively Self-Improving Agent for Deep Research
AREX is built around repeated constraint checks, using partial verification to decide what the agent researches next.
The paper frames deep research as easier to verify in pieces than to solve in one pass. AREX alternates between gathering evidence, drafting a provisional answer, auditing it claim by claim, and sending the agent back after unresolved points. The authors also describe a learned context-update tool meant to preserve verified evidence and open constraints over long runs. They report results from 4B and 122B-A10B models that outperform comparable baselines across several research, reasoning, and tool-use benchmarks. ArXiv · AI/CL/LG's note
The paper frames deep research as easier to verify in pieces than to solve in one pass. AREX alternates between gathering evidence, drafting a provisional answer, auditing it claim by claim, and sending the agent back after unresolved points. The authors also describe a learned context-update tool meant to preserve verified evidence and open constraints over long runs. They report results from 4B and 122B-A10B models that outperform comparable baselines across several research, reasoning, and tool-use benchmarks. ArXiv · AI/CL/LG's note
score 5