cfli/AREX-2
AREX-2 is a 27 billion parameter multimodal agent model developed by the Beijing Academy of Artificial Intelligence (BAAI), built on a Dense Qwen3.8-compatible architecture. It specializes in long-horizon self-improvement, iteratively refining solutions by processing feedback such as scores, logs, and errors. This model excels at complex tasks requiring sustained reasoning, including algorithmic programming, machine-learning engineering, and deep research, demonstrating strong performance across these domains.
Loading preview...
AREX-2: Long-Horizon Self-Improving Agent
AREX-2 is a 27 billion parameter multimodal agent model from the Beijing Academy of Artificial Intelligence (BAAI), designed for long-horizon tasks. Built on a Dense Qwen3.8-compatible architecture, it features a substantial context length of 262,144 tokens.
Key Capabilities
- Long-Horizon Self-Improvement: AREX-2 can refine solutions over multiple test-time rounds by proposing, measuring, reflecting, and revising.
- Feedback-Driven Reflection: The model effectively uses various forms of feedback, including scores, logs, errors, and timings, to guide its iterative improvements.
- Cross-Domain Performance: Trained on machine-learning and algorithmic-programming tasks, its self-improvement capabilities transfer effectively to deep research problems.
- Sustained Reasoning: It maintains productive iteration even as task complexity and budget increase.
Evaluation Highlights
AREX-2 demonstrates strong performance across various benchmarks:
- Coding & ML Engineering: Achieves 70.7 on Frontier-CS and 81.8 on MLE-Lite, outperforming many larger open-weight models.
- General Agentic Reasoning & Deep Research: Scores 84.0 on BrowseComp, 52.6 on HLE, 92.2 on GAIA, and 93.8 on DeepSearchQA, competing favorably with frontier models in its size class.
Good for
- Research into long-horizon agents and iterative problem-solving.
- Machine-learning engineering and algorithmic coding tasks.
- Tool-augmented deep research applications.