近日，前腾讯 AI Lab R4 部门合伙人、“王者绝悟”技术负责人叶德珩（Deheng Ye）在个人主页更新了职业信息，确认已加入字节跳动 Seed 团队，担任研究员，从事基础模型（Foundation Models）相关研究。
叶德珩在托管于 GitHub Pages 的个人主页中写道：
I am a Researcher at ByteDance Seed working on foundation models.
即“我目前是字节跳动 Seed 团队的一名研究员，从事基础模型研究”。
页面列出的工作履历显示，他于 2017 年至 2026 年供职腾讯，2026 年起加入字节跳动。
在项目经历中，除腾讯时期的“绝悟”系列外，还新增了 SeedXxx 模型，并标注“P&C for now”，意味着相关信息目前尚未公开。
页面还提到了生成媒体方向的智能体系统，包括 ui-voyager、omni-reasoner 和 proact。
他同时留下了字节跳动企业邮箱用于招募人才。
今年 3 月 22 日，雷峰网在《腾讯 AI Lab 撤销背后：俞栋离职是最大导火索》一文中披露，叶德珩已于 2 月离开腾讯。当时，他所在的 R4 团队成立尚不足半年。其离职后，团队被调整至薄列峰负责的多模态体系。
这意味着，在腾讯正式宣布撤销 AI Lab 之前，叶德珩就已经离开了这家工作近九年的公司。
从“王者绝悟”到大模型，腾讯近九年的技术积累
叶德珩并非传统意义上仅从事大语言模型研究的科学家。
他的研究经历横跨软件工程、深度强化学习、多智能体系统与基础模型，尤其以腾讯时期的游戏 AI 研究受到关注。
公开履历显示，他于 2016 年在新加坡南洋理工大学取得计算机科学博士学位，随后于 2016 年至 2017 年在微软从事博士后研究，方向涉及代码理解与生成。
2017 年加入腾讯后，他逐渐将研究重点转向游戏 AI、强化学习和大规模机器学习系统。
2021 年，腾讯曾在官方人物报道中专门介绍叶德珩。当时，他担任腾讯技术工程事业群悟空 AI 团队负责人，带领团队将机器学习技术应用到《王者荣耀》等复杂游戏场景。
在腾讯官方介绍中，叶德珩负责复杂问题的关键技术方案研发，并带领团队建设端到端 AI 平台。
其参与的“绝悟”系列项目，则是腾讯在强化学习领域的重要研究成果之一。
相较于围棋等回合制环境，《王者荣耀》具有连续决策、高维动作空间、多角色协作、实时对抗等特点。智能体不仅要完成局部操作，还需要在动态环境中处理长期策略与团队协作问题。
2020 年，叶德珩等人的研究成果《Mastering Complex Control in MOBA Games with Deep Reinforcement Learning》介绍了在复杂 MOBA 游戏中应用深度强化学习的方法。相关工作涉及动作空间优化、策略学习与高性能训练系统。
据当时研究介绍，AI 智能体在测试中能够使用不同英雄战胜顶尖职业选手。
此后，团队进一步探索分层强化学习。
2022 年，叶德珩作为作者之一参与的论文《JueWu-MC: Playing Minecraft with Sample-efficient Hierarchical Reinforcement Learning》发表于 IJCAI，研究如何让 AI 在《我的世界》这类开放环境中高效学习复杂任务。
这类研究与今天的 Agent 有一定的技术关联：智能体需要理解环境、规划行动、完成子任务，并根据反馈不断调整策略。
但两者不能简单等同。游戏强化学习所积累的技术经验，是否以及如何迁移到当下的基础模型训练，需要结合具体项目判断。
叶德珩后来也将研究范围扩展至大语言模型与多智能体协作。
2024 年的论文《More Agents Is All You Need》探索通过增加参与采样和投票的智能体数量，提高大语言模型在部分任务上的表现。
从游戏 AI 到多智能体系统，再到基础模型，他的研究方向经历了逐步扩展，而非突然转型。
在腾讯内部，叶德珩的技术与管理职级也曾多次晋升。
他在最新个人主页中写道，自己曾在 2025 年之前成为腾讯 AI 体系内最年轻的 Leader、Director 和 Research Partner。
据雷峰网此前报道，腾讯 AI Lab 曾设置 R1 至 R4 四个部门，并采用合伙人制。叶德珩担任 R4 部门合伙人，其他部门合伙人包括俞栋、胡瀚和冯佳时等研究人员。
腾讯 AI Lab 撤销前后，多位技术骨干离开
叶德珩的离职发生在腾讯 AI 组织体系重大调整期间。
2026 年 3 月 20 日，腾讯内部通知撤销成立近十年的 AI Lab，部分成员调整至大语言模型部，加入混元团队，并向首席 AI 科学家姚顺雨汇报。产学研合作中心则得到保留。
腾讯公司副总裁蒋杰不再兼管 AI Lab，相关工作移交至新的组织体系。在此之前，腾讯 AI Lab 已发生过多次人才和团队调整。
雷峰网 3 月 22 日的报道指出，AI Lab 原副主任俞栋离开后，其语音研究团队被划入薄列峰负责的多模态部门。叶德珩则在今年 2 月离职，R4 团队随后也被调整至相关体系。
根据第一财经报道，AI Lab 调整涉及原有研究组织的重新整合：部分人员进入姚顺雨负责的混元体系，另有团队保留在技术工程事业群下的相关组织中。
从组织架构来看，腾讯试图将过去相对独立的基础研究力量，与大模型研发和工程体系进一步结合。
此前，腾讯已经在 2025 年 12 月围绕姚顺雨重新调整 AI 组织架构，设立 AI Infra 部门，并强化大语言模型研发力量。3 月撤销 AI Lab，是这一系列组织调整的延续。
但叶德珩究竟为何选择离职，目前本人并未亲自公开解释。
此次新动态可以看到，他的研究方向和此前略有不同。从其过往研究经历来看，强化学习、多智能体协作等技术积累，与当前基础模型及 Agent 的发展方向存在一定契合，但其在 Seed 团队承担的具体职责仍有待进一步更新。
从这个角度看，叶德珩从游戏 AI 转向基础模型也不那么令人意外。
他的经历也反映出一个更广泛的行业变化：大模型竞争正在吸引来自传统 AI 研究实验室、强化学习、机器人、游戏智能体等不同技术领域的人才。
过去，这些研究人员可能分布在独立实验室和垂直业务团队；如今，越来越多的研究方向开始与基础模型研发产生交汇。
对于字节 Seed 而言，叶德珩的加入意味着团队新增了一位具有强化学习和多智能体研究经历的资深研究人员。
从“王者绝悟”到基础大模型，叶德珩能否在字节 Seed 延续此前的技术积累，带来新的研究成果，还是值得期待的。
参考链接：
https://www.tencent.com/zh-cn/tencenttalent-a-spirit-of-craftmanship-brings-technology-to-life/
https://github.com/YeDeheng/yedeheng.github.io/blob/master/index.html
https://arxiv.org/pdf/2402.05120
https://ojs.aaai.org/index.php/AAAI/article/view/6144
https://www.aminer.cn/profile/deheng-ye/562cfdd945cedb3398d212a1
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
In recent days, Deheng Ye, the former Tencent AI Lab R4 Sector Partner, technical director of the King's Consciousness, updated his career information on his home page, confirming his membership in the byte beat Seed team as a researcher and conducting research on the Foundation Models.
In a personal homepage hosted by GitHub Pages, Yedwin writes:
I am a Researcher at ByteDance Seed working on foundation models.
"I am currently a researcher of the byte beat Seed team, doing basic modelling."
The job resumes listed on the page show that he worked for Tencent from 2017 to 2026, and that he joined byte since 2026.
In the course of the project, in addition to the Tencent period's “discovery” series, the SeedXxx model was added with the label “P&C for now” meaning that the relevant information is not currently available.
The AI agent system, which generates media orientation, is also mentioned, including ui-voyager, amni-reasner and case.
At the same time, he left the byte jumper business mailbox for recruitment.
On March 22, this year, Le Peak Network revealed in the book Tencent AI Lab Uncovered: RYU's departure was the biggest fuse, and Yedwin left Tencent in February. At the time, his R4 team was less than six months old. After their separation, the team was repositioned to the multi-modular system of responsibility at the thin peak.
This means that before Tencent officially announced the withdrawal of AI Lab, Yedwin left the company for almost nine years.
From King's Awareness to LLM, Tencent, nearly nine years of technological accumulation
Yedwin is also a non-traditional scientist engaged only in large-language modelling.
His research experience spans software engineering, intensive in-depth learning, multi-AI agent systems and basic models, especially with the Tencent game AI study.
The public curriculum vitae shows that he obtained a Ph.D. in Computer Science from Nanyang Polytechnic University in Singapore in 2016, followed by a postdoctoral study in Microsoft from 2016 to 2017 in the direction of code understanding and generation.
After joining Tencent in 2017, he gradually shifted the focus of his research to play AI, intensive learning and large machine learning systems.
In 2021, Tencent gave a special presentation on Yed-in in official figures. At that time, he was the head of the Tencent Technological Engineering Enlightenment AI team, leading the team in applying machine learning techniques to complex games such as Kings Glory.
In the official presentation of Tencent, Yedwin was responsible for developing key technical programmes on complex issues and leading team-building to the end-to-end AI platform.
Its participation in the “Desperation” series is one of the key research findings of Tencent in the field of intensive learning.
The Kings Glory has the characteristics of continuous decision-making, high-dimensional action space, multi-role collaboration, real-time confrontation, and so on, compared to a round-making environment such as chess. AI agent not only performs local operations but also addresses long-term strategy and teamwork in a dynamic environment.
In 2020, Mastering Complex Control in MoBA Games with Deep Reinforcement Learning introduced the application of intensive learning in complex MoBA games. The work involves action space optimization, strategic learning and high performance training systems.
According to the study at that time, AI AI agent was able to use different heroes to win over top professional players.
Since then, teams have further explored tiered and intensive learning.
In 2022, the paper entitled " JueWu-MC: Playing Minecraft with Sample-efficient Hierarchical Renewal Learning " was published in IJCAI to study how to enable AI to learn complex tasks efficiently in an open environment such as My World.
Such research has some technical relevance to today's Agent: AI agent needs to understand the environment, plan action, complete subtask, and adjust the strategy in the light of feedback.
But the two cannot be simply equated. The technical experience gained by the enhanced learning of games, whether and how to migrate to current basic model training needs to be judged in the context of specific projects.
Yedwin also later extended the scope of the study to the Big Language Model in collaboration with Multi AI agent.
The 2024 paper, More Ages Is All You Need, explores ways to improve the performance of large language models in part of the mission by increasing the number of AI agents involved in sampling and voting.
From the game AI to the multi-AI agent system, to the base model, his research direction has been gradually expanded rather than suddenly transformed.
Within Tencent, the technical and managerial levels of Yedhwan were also promoted on several occasions.
In his latest home page, he wrote that he had become the youngest Leader, Director and Research Partner in the Tencent AI system before 2025.
Tencent AI Lab has previously reported that four departments, from R1 to R4, have been set up by Tencent AI Lab and have been partnered. Yedwin is a partner in the R4 sector, and other sector partners include researchers such as RYO, Huan and Feng Jia time.
Before and after Tencent AI Lab is withdrawn, multiple technical cadres leave.
The departure of Yedwin occurred during a period of significant change in the Tencent AI organizational system.
On March 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 26, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 30, 30, 28, 28, 30, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 28, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, The Cooperative Centre for Productive Studies and Research was retained.
Jiang Jie, Vice-President of Tencent, no longer runs AI Lab and has been transferred to the new organizational system. Prior to this, Tencent AI Lab has experienced numerous talent and team adjustments.
On March 22, a report by Ray Peak reported that after AI Lab’s former deputy director, Zhao, had left, his voice research team was assigned to the multi-modular unit in charge of the Bottle Peak. Yedwin left in February of this year, and the R4 team was subsequently restructured into the system.
According to the first financial report, AI Lab has been re-organised to involve the reintegration of former research organizations: some have entered the mixed system responsible for Yao Sun Rain, while others remain in the relevant organizations within the technical engineering community.
From an organizational point of view, Tencent seeks to further integrate the former relatively independent basic research force with the LLM R & D and engineering system.
Previously, in December 2025, Tencent had reorganised the AI organizational structure around Yao Sun Rain, set up the AI Infra department and strengthened the development of large language models. The withdrawal of AI Lab in March is a continuation of this series of organizational adjustments.
However, the reason why Yedwin chose to leave his post was not publicly explained in person.
As can be seen from this new development, the direction of his research is slightly different from what it was before. From its previous research experience, the accumulation of technologies such as enhanced learning, multi-AI agent collaborations is in line with the current basic model and Agent ' s direction, but its specific responsibilities in the Seed team still need to be further updated.
It's not surprising from that point of view that Yedwin moved from the game AI to the base model.
His experience also reflects a broader industry change: LLM competition is attracting talent from different technical fields such as traditional AI research laboratories, intensive learning, robots, and games AI agent.
In the past, these researchers may have been located in independent laboratories and vertical teams; more and more research directions are now beginning to intersect with basic model development.
For byte Seed, Yedwin ' s addition meant that the team added a senior researcher with enhanced learning and multi-AI agent research experience.
From the King's Awareness to the Foundation LLM, it is still to be hoped that Yed-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e-e
References:
https://www.tencent.com/zh-cn/tencenttalent-a-spirit-of-craftmanship-brings-technology-to-life/
https://github.com/YeDeheng/yedeheng.github.io/blob/master/index.html
https://arxiv.org/pdf/2402.05120
https://ojs.aaai.org/index.php/AAAI/article/view/6144
https://www.aminer.cn/profile/deheng-ye/562cfdd945cedb3398d212a1
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.