DeepSeek-R1超级外挂！“人类最后的考试”首次突破30分，上海交大等开源方案碾压OpenAI、谷歌

Their open-source tool-enhanced reasoning agent, X-Master, utilizes the DeepSeek-R1-0528 model and introduces a novel multi-agent workflow to enhance problem-solving capabilities.

Key Points

• X-Master scored 32.1, the first system to exceed 30 on HLE.
• The model uses DeepSeek-R1-0528 and features a novel multi-agent workflow.
• X-Masters outperformed existing models in all tested categories.
• The approach combines tool-enhanced reasoning with structured exploration.
• X-Master achieved 67.4% accuracy in complex biology tasks.