On March 12, Yu Bowen, previously the head of post-training for Alibaba’s Qwen large model at Tongyi Lab, officially joined ByteDance. He will lead post-training efforts for the visual model and multimodal interaction team within ByteDance’s Seed initiative.

This move came just five days after Yu Bowen announced his resignation on March 7. Coincidentally, his departure aligned with that of Lin Junyang, another prominent member of Alibaba’s Qwen team. Furthermore, Yu Bowen is not the first key figure from Alibaba’s Qwen team to transition to ByteDance. In July 2024, Zhou Chang, the former technical lead of Tongyi Qwen, was also recruited by ByteDance with a lucrative multi-million annual salary.

The back-to-back departures of two technical leaders to ByteDance raise questions about ByteDance’s underlying AI strategy. Additionally, it highlights why post-training experts have become a highly sought-after and scarce resource in the competitive landscape of large models.

From ‘Alibaba Star’ to Post-Training Lead

Public records indicate that Yu Bowen completed his undergraduate studies at Central South University and later pursued graduate studies at the Institute of Information Engineering, Chinese Academy of Sciences. He earned his Ph.D. from the University of the Chinese Academy of Sciences in 2022, focusing on natural language processing and information extraction. During his academic career, he published multiple papers at top international conferences such as ACL and EMNLP. He introduced an innovative method for transforming information extraction tasks into graph-structured problems. His outstanding academic achievements earned him the President’s Award from the Chinese Academy of Sciences.

After earning his Ph.D. in 2022, Yu Bowen joined Alibaba’s DAMO Academy through the ‘Alibaba Star’ program—Alibaba Group’s premier campus recruitment initiative—as an algorithm expert (P7). From the beginning, he played a pivotal role in the early training and development of the Tongyi Qwen large model, quickly becoming a core member of the Qwen team and eventually leading post-training efforts.

As the post-training lead, Yu Bowen made significant contributions to the ‘alignment’ and ‘fine-tuning’ of large models. He spearheaded the development of the Qwen series Chat models, utilizing techniques such as Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), and Direct Preference Optimization (DPO) to adapt general-purpose large models into dialogue models aligned with human values and usage patterns.

For models of varying scales, he designed a ‘four-stage evolutionary theory’ and ‘knowledge distillation’ strategies, enabling the Qwen models to excel in tasks like long-text generation, complex reasoning, and multimodal understanding.

Yu Bowen’s career progression, from ‘Alibaba Star’ to post-training lead, reflects the broader trajectory of AI talent development in China and has positioned him as a highly coveted professional in the industry.

Catalyst for Departure

Industry observers speculate that Yu Bowen’s departure was prompted by organizational restructuring at Alibaba.

In early March, Alibaba’s Tongyi Lab initiated a restructuring plan to transition the vertically integrated Qwen team into multiple horizontally aligned modules, including pre-training, post-training, text, and multimodal teams. This restructuring significantly narrowed Yu Bowen’s management scope and clashed with his technical philosophy of ‘deep coupling between pre-training and post-training.’

Simultaneously, increased commercialization pressure from Alibaba’s senior management on the Qwen team exacerbated internal divisions. Insiders revealed that Alibaba set Daily Active Users (DAU) as a key performance indicator for the open-source team responsible for foundational large model research. This forced the team to allocate substantial resources to activities like promotional campaigns, simplifying model functionalities, and integrating with Alibaba’s ecosystem consumption scenarios.

This shift from a technology-first to a business-first approach directly conflicted with the ‘extreme open-source, zero commercial cost’ philosophy advocated by core technical personnel like Lin Junyang and Yu Bowen. Lin Junyang had repeatedly emphasized internally the need for closer collaboration among pre-training, post-training, and even Infra and training teams.

ByteDance’s Strategy

Yu Bowen has joined ByteDance’s Seed team, a central hub for AI research and development. The team is currently led by Dr. Wu Yonghui, former Vice President of Research at Google DeepMind, who contributed to the development of the Gemini large model. Wu Yonghui officially joined ByteDance in February 2025 and assumed leadership of the Seed team, reporting directly to ByteDance CEO Liang Rubo.

The Seed department’s research encompasses large language models (LLMs), speech, vision, world models, infrastructure, AI Infra, and next-generation AI interactions. Its developed Doubao large model has been deployed in over 50 scenarios. The team focuses on advancing multimodal technology, having iteratively released core products such as the Seed 2.0 series foundational models, Seedance 2.0 video generation model, and Seed3D 1.0 3D generation model.

Yu Bowen’s arrival will undoubtedly bolster ByteDance’s post-training capabilities in visual and multimodal interactions. His technical expertise in supervised fine-tuning, reinforcement learning, and direct preference optimization, honed during the development of the Qwen series Chat models, aligns precisely with ByteDance’s needs for multimodal model alignment and fine-tuning.

Notably, the ‘multimodal interaction and world models’ department that Yu Bowen will join is headed by Zhou Chang, the former technical lead of Alibaba’s Qwen. This suggests that ByteDance is systematically assembling a multimodal technology team comprised of core members from Alibaba’s Qwen.

As competition among large models intensifies, the value of post-training experts is being reevaluated and contested.

While pre-training determines a model’s knowledge breadth and foundational capabilities, post-training shapes its practical usability, safety, and user experience. An exceptional post-training expert can transform a massive foundational model into a truly functional product through sophisticated algorithm design and engineering implementation.

ByteDance’s strategic recruitment of key post-training personnel from Alibaba’s Qwen team reflects its clear vision in AI. The company aims not only to enhance its foundational model capabilities but also to establish a competitive edge in model practicality and productization, in line with ByteDance’s product-driven ethos.

Epilogue

In the relentless AI race, the quest for top talent remains constant. Companies that can offer technical ideals and growth opportunities to such talent will secure a long-term competitive advantage.

As stated in The Three-Body Problem, ‘Give civilization to the years, not years to civilization.’ In the AI technological revolution, the truly scarce resources are not computational power and data but the post-training experts who can ‘imbue models with soul.’ As large models evolve from ‘brute force miracles’ to ‘meticulous craftsmanship,’ the decisive factor is no longer parameter scale but the ‘invisible hand’ that enables AI to truly comprehend human intentions.