After the vice president departed, another core T10-level leader at Baidu Wenxin has left the company.

Another key member has left the Wenxin team.

On November 4, multiple media outlets broke the news that Liu Jing, Distinguished Architect at T10 level and head of post-training for Baidu's Wenxin foundational large model, has left Baidu. The news spread rapidly in the AI circle only one week after Wu Tian, Baidu's vice president, confirmed her departure from the company.

Public resumes show that Liu Jing received his bachelor's degree from Xidian University, and his master's and doctoral degrees from Harbin Institute of Technology. Before joining Baidu, he worked as a researcher at Microsoft Research Asia (MSRA), focusing on information retrieval, question answering and knowledge bases.

He joined Baidu in 2017, starting from the question answering and information retrieval line of the Natural Language Processing Department. He led the development of DuReader 2.0, a Chinese machine reading comprehension dataset, and KT-NET, a technology that directly embeds knowledge into pre-trained representations.

Around 2019, when Baidu was launching its knowledge-enhanced large model ERNIE, Liu Jing was one of the early core members of this technical route. Academically, he has published more than 20 papers at conferences including ACL and SIGIR, with over 1,600 citations on Google Scholar. He has won first place in many international evaluations such as MRQA, MSMARCO and NTCIR, and also served as the area chair for the question answering track at ACL and AACL.

Before his resignation this time, his title was head of post-training for Wenxin foundational large model, who was fully in charge of the entire end-to-end chain including data construction, algorithm optimization and evaluation system, and led the R&D and large-scale implementation of general agents, code agents and deep search.

The fields he has long been deeply engaged in — large model alignment, reinforcement learning in the inference stage, intermediate pre-training, and agent reinforcement learning — are exactly the most critical segments for current large models to evolve from "being able to chat" to "being capable of performing practical tasks".

As for his next employer, there are two versions circulating on the Internet at present.

One version says he will go to Zhongguancun Institute in Beijing, a new type of research institution led by Liu Tieyan that focuses on the coordinated development of government, industry, academia, research and finance; the other says he has joined a company called SmartHey to continue working on large model alignment and agents. Neither of the two claims has official sources, and Liu Jing himself has not made any public statement. Some media first pointed out on November 3 that he "may join Zhongguancun Institute", but the wording used was "may".

As of press time, Baidu has not made any public response to Liu Jing's resignation.

When the last similar news came out, the official caliber was totally different. When Wu Tian left Baidu on October 29, a Baidu public relations manager confirmed on WeChat Moments that "Wu Tian has ended her tenure at Baidu", and added that "Baidu's technical work is still moving forward". That sentence seemed to be aimed at stabilizing the situation, but also seemed to admit that something was loosening.

Earlier, on April 28, Baidu sent a company-wide letter announcing that it would cancel the long-used T, P, E, Band and M job level labels starting from May 1, integrate the professional track and management track, and unify them into digital job levels ranging from 5 to 12. In May, Baidu also established the Model Committee (BMC). The organization is being flattened, and decision-making power is being delegated to front-line teams, but employees are also leaving in batches.

The timing of the incident is thought-provoking.

The day after the news of Liu Jing's resignation was exposed, on November 5, Baidu just open-sourced Wenxin 5.5 Flash Base — which adopts the MoE architecture, has a total of 280 billion parameters, activates 14 billion parameters per Token, uses 33.5 trillion Tokens for pre-training, and supports a maximum context length of 1 million Tokens during the fine-tuning and inference stages. The Baidu World 2026 conference is about to be held on November 12.

New models are being released, conferences are upcoming, but veteran core employees are leaving.

The post-training link is responsible for how to align the model after pre-training, make it follow instructions, and use reinforcement learning to maximize its capabilities. Unlike pre-training which can boast of trillions of parameters, post-training is the critical threshold that determines whether a model can be truly implemented in real scenarios and evolve into an intelligent agent. Baidu has not given any explanation for the departure of this T10 level talent, and no public successor has appeared for the time being.

Wenxin is still iterating, but the core team members on this key line have been replaced batch after batch.

Source: Xinghe Business Observation