BAAI / AREX-Base

Towards a Recursively Self-Improving Agent for Deep Research

Introduction

AREX is a family of deep research agents developed by the Beijing Academy of Artificial Intelligence (BAAI). It is designed for long-horizon tasks in which an agent must search across sources, assemble candidate answers, verify multiple constraints, and revise its research plan when the available evidence is incomplete.

AREX follows a recursive self-improvement framework consisting of an inner research loop and an outer self-improvement loop. The inner loop conducts deep research, while the outer loop evaluates and improves the research process.

This repository contains AREX-Base , the larger model in the AREX family. It is built on Qwen3.5-122B-A10B and uses a Mixture-of-Experts architecture with 122B total parameters and 10B activated parameters.

• Architecture: Mixture-of-Experts
• Parameters: 122B
• Context length: 262,144 tokens

Key Features

• Recursive self-improvement: evaluates provisional answers and chooses whether to accept, refine, or restart the research trajectory.
• Verification-guided research: converts unresolved answer constraints into focused follow-up searches.
• Autonomous context updates: retains verified findings, source identifiers, rejected candidates, open constraints, and the next research plan.
• Long-horizon tool use: supports multi-round search, browsing, evidence integration, and answer construction.
• Evidence-grounded answers: maintains supporting evidence and answer-level confidence throughout the research process.

Model Family

Method

AREX organizes deep research as two interacting loops supported by autonomous context management.

1. Inner research loop. The agent searches, reads, integrates evidence, tracks candidates, and produces a provisional answer with supporting evidence and an answer-level confidence score.
2. Outer self-improvement loop. The agent checks the provisional answer against the original constraints. High-confidence answers are accepted. Recoverable trajectories are refined around unresolved claims, while uninformative trajectories can be restarted.
3. Autonomous context update. The model invokes update_context to refresh its research state around verified findings, current and rejected candidates, unresolved constraints, source validity, and the next research plan.

This design preserves useful work across iterations and reduces repeated exploration of paths that have already been rejected.

Evaluation

AREX is evaluated through a unified long-horizon search-agent interface with search , visit , update_context , and finish tools. HLE with tools additionally provides a Python tool.