ByteDance Seed recently announced the establishment of a "Seed Model-Alignment & Preparedness" department, led by Yan Lin and reporting to Seed head Wu Yonghui. The former Responsible AI team and some members of the Chat team will be transferred into the new department, further strengthening investment in the field of model safety and alignment.
The new department will focus on improving the safety capabilities of agent models and guarding against potential risks posed by frontier capabilities such as chemical, biological, radiological and nuclear (CBRN) threats. Its work will cover the formulation of model specifications, the building of safety evaluation systems, alignment training, the construction of protection systems and risk monitoring, and it will strengthen open collaboration with external research institutions and third-party evaluation organizations.
It is reported that, as large model capabilities and autonomy continue to grow, safety issues such as model behavior deviating from expectations, agent misuse and cyberattacks are receiving more attention. Previously, safety and alignment work at domestic large model teams was mostly incorporated into companies' overall security and compliance systems. This adjustment means that Seed will build safety and alignment as an independent function. According to relevant information, ByteDance Seed is the first frontier AI laboratory in China to establish a separate safety and alignment department.
Within Seed, it is believed that safety and alignment not only concern risk prevention, but also affect whether models can achieve large-scale application in complex scenarios. The new department will be committed to ensuring that frontier models effectively execute tasks under safety specifications and authorization constraints, continuously reducing the risks of model misuse and behavioral deviation, and improving the safety system for unknown risks.