BAAI Releases AREX: 10B-Activation Agent Rivals GPT-5 in Autonomous Research
新智元 · wechat · 2026-08-10
The Beijing Academy of Artificial Intelligence (BAAI) has released AREX, an autonomous research agent framework, along with open-source model weights. The framework centers on solving the "discovery-verification asymmetry" via a dual-loop recursive system: an inner Research Loop (searching and drafting tentative answers) and an outer Self-Improvement Loop (auditing constraints and identifying gaps).
Core Mechanics & Training:
- Autonomous Context Update (ACU): The model actively maintains a "research progress sheet" instead of raw history, preserving verified facts and excluding dead ends to prevent context loss.
- Training Strategy: AREX uses verifiable tasks generated from known answers. It employs staged training and step-aware reinforcement learning, focusing on critical turning points like evidence discovery and search redirection.
Benchmark Performance:
- High Parameter Efficiency: AREX-Base has 122B total parameters but only activates 10B. It achieves frontier-level performance across six benchmarks, including BrowseComp and DeepSearchQA.
- Beating Giants: On WideSearch-en, it outperforms closed-source flagships like GPT-5.4 and Claude Opus-4.6. It also comprehensively beats the much larger Qwen3.5-397B on multiple metrics.
Additionally, BAAI launched the AREX Research Platform, packaging this capability into a daily research and tracking tool for academic and industrial users.
More from Models
- Jeff Dean Departs, DeepMind Overhauled, Gemini 3.5 Pro Cancelled: SemiAnalysis — koltregaskes · 2026-08-10
- Prompt Over Vision: ChatGPT Hallucinates Image Details Based on Text Cues — Red-Pony · 2026-08-10
- Speculation: Flash-0731 Might Be a Blinded Vision Model — teortaxesTex · 2026-08-10
- Sarvam AI Launches 105B Model, Claims to Beat GPT at 20% Cost — itsOmSarraf_ · 2026-08-10
- Gemini 3 Sets New SOTA on ARC-AGI-2, Deep Think Hits 45% — typewriters · 2026-08-10
- Pinterest Earnings: Open Models Cost Under 8% of Closed Alternatives — juliey4 · 2026-08-10