ByteDance restructures Seed AI team to centralize data, RL and product training
ByteDance reorganized its Seed AI unit in August, creating four departments for pre-training data, reinforcement learning and product post-training for workplace and chat. The shift signals a task-oriented approach as AI competition moves to agentic and omni-modal development.
The new departments are Pretrain Data, Horizon RL, Product Posttrain-Work and Product Posttrain-Chat, led respectively by Li Chenggang, Tang Shengyu, Qin Yujia and Zhu Wenjia, all reporting to Wu Yonghui. Seed said in an internal announcement that the move was meant to merge similar work and reduce overlap.
Pretrain Data brings together pre-training data teams that were previously split across text, code, vision and speech. It will now provide data for ByteDance's omni-modal and very large models. Li Chenggang, who built ByteDance's search system and led search engineering for Toutiao and TikTok, is in charge of the unit.
Horizon RL, led by Tang Shengyu, consolidates post-training teams for reasoning, visual understanding and other directions. Its three groups are RL scaling, reasoning and visual reinforcement learning, headed by Yue Yu, Wang Mingxuan and Wu Youbin respectively. Seed has published reinforcement-learning systems DAPO and VAPO, and the new unit is tasked with improving the basic intelligence ceiling of the models rather than fine-tuning a single output.
Product-side work has been split by user task rather than by model capability. Product Posttrain-Work, headed by Qin Yujia, serves workplace and agentic scenarios, including task modes in Doubao and Dola. Under it are a GUI branch led by Wu Zhiyong and an expanded branch built from the former Agent office team. Product Posttrain-Chat is the renamed Application team still led by Zhu Wenjia; it focuses on consumer dialogue, search-based Q&A, personalization and inference cost control. Yan Lin, former head of Doubao's alignment team, manages its Doubao and Dola Chat branches.
The new structure creates some unusual reporting lines across departments. For example, visual reinforcement learning sits inside Horizon RL under Wu Youbin, but Wu Youbin also reports to Lin Yi, head of visual pre-training in Pretrain Data, on the visual track. Tang Shengyu also supervises code pre-training, according to Leiphone.com.
The reorganization is part of a broader trend in the AI industry. Tencent last month merged its Hunyuan large-language-model team and multimodal-model team into one basic-model department under Yao Shunyu. Meta has regrouped its superintelligence lab around model, research, product and infrastructure. As frontier models become omni-modal and agents take on tool execution in products, ByteDance is abandoning an organization built around text, code and vision silos in favor of data, reinforcement learning and end-user tasks. Whether the new framework can run smoothly in practice remains the next question.