2026 年夏季项目将于 6 月至 8 月举行,这将是 MATS 迄今规模最大的一届,共有 120 位学员和 100 位导师。学员将与导师或机构研究团队合作,围绕研究项目开展整个夏季的工作,合作方包括 Anthropic 对齐科学团队、UK AISI、Redwood Research、ARC 和 LawZero。部分学员将获邀参加为期 6 个月以上的延长期,继续这项合作。
目前已截止申请。

申请与招生流程中的关键阶段
申请通常会在项目开始前几个月开放。申请者需完成多阶段招生流程,首先提交通用申请。根据所选研究方向、研究小组和导师,申请者可能还需参加面试、工作测试、编程评估或提交写作样本,之后才会作出最终录取决定。
主项目为期 10 至 12 周,是全职研究项目。学员与一位或多位导师密切合作,独立开展研究,同时参加工作坊、讲座、导师答疑时段,并融入更广泛的 MATS 社群。学员与导师共同确定研究方向,并在项目期间逐步提升独立开展研究的能力。
延伸阶段通常在主项目结束约两周后开始。主项目中展现出较强研究潜力的学员可申请为期 6 至 12 个月的资助延伸项目。延伸阶段的学员在持续获得导师指导和支持的同时,继续推进独立研究,通常在 MATS 办公室或其他获准的研究地点工作。最近几期中,约 80% 申请延伸阶段的学员获得录取。
MATS 致力于帮助研究人员加速发展,走向以下方向:
MATS 校友后来发表了安全研究成果、加入了 Anthropic 和 MIRI 等对齐研究组织,也创办了对齐研究实验室。更多校友信息请点击这里。
每个 MATS 研究小组都由学员和导师组成,围绕共同的研究议程开展工作。各小组的方法和重点各不相同,涵盖可解释性、控制、评估、治理、网络安全和智能体基础等主题。
发生灾难时,政府可能不得不成为最终保险人,无论是为某个实验室兜底,还是为整个社会承担风险。(不止我这么认为,例如这篇文章也有类似观点。)设计得当的联邦保险兜底机制,可以让造成灾难性后果的负外部成本由相关方承担;设计不当,则只会变成对 AI 公司的补贴。我希望设计出合理的方案,提前做好准备。
我鼓励具备机制设计(也称逆向博弈论)专长的人申请;没有这方面经验也不必因此却步。
通常每周安排一小时会议,提供整体方向指导。我会通过 Slack 保持联系,简单问题或概念讨论(不包括代码调试)通常会在一天内回复。两次会议之间,我们会异步讨论论文结构、实验设计和研究结果。学员遇到困难或想一起构思时,也可以在 Slack 上联系我,临时安排通话。
根据项目需要,我也可能提供写作支持。
如果有意参与技术论文项目,申请者必须:
所有申请者:
优先考虑:
有帮助的条件:
不太适合的情况:
对于技术类项目,具体范围可能会根据学员的专长自然收窄。我会提出核心难题,第一周里再与学员一起明确需要回答的理论问题,以及需要开展哪些实证调查。
对于非技术类项目,我会介绍几个备选方向,学员先尝试自己感兴趣的项目一周,第二周再共同确定最终方向。
本项目关注语言模型如何形成支撑泛化能力的内部表征,例如关于角色、目标或训练数据组成部分的表征。
每周安排一小时会议,并在 Slack 话题串中进行异步讨论;如有需要,也可以临时安排额外会议。
必备条件:
优先考虑:
我们会在项目开始时一起梳理可选方向,学员也可以提出其他想法。学员应在最初一两周探索不同项目,并在第二周确定研究方向。
我们研究奇异学习理论(SLT)在 AI 安全中的应用,重点关注可解释性与对齐。理想的学员应具备数学、物理、计算机科学或生物学方面的扎实技术背景,也愿意亲自开展机器学习实验。无需深入掌握 SLT,但了解一些基础会有所帮助。
团队每周会与两位导师一起开会。此外,你每两周会与至少一位导师进行一次一对一交流。团队通过内部 Discord 服务器进行异步沟通。我们也希望你根据需要,主动安排与团队成员结对编程或共同调试。
遇到研究上的困难时,我们会提供帮助,包括在会议之外。
If you're interested in working on more of the empirical side, you should have prior experience with ML engineering (at least at the level of a program like ARENA) and prior research experience (potentially in a field outside of ML). A bonus would be prior familiarity with designing and running ML experiments or research specifically in AI safety.
If you're interested in working on more of the theoretical side, you should have prior research experience in a relevant field like mathematics, theoretical physics, or theoretical computer science.
Please make sure that your background and interests are clearly described in your application. By default, we'll be looking for evidence of research ability in the form of publications.
We do not expect you to already be aware of SLT, but if you pass the first round, please prepare by conducting some background reading (see: timaeus.co/learn).
导师会与学员讨论项目构想,并提供几个方向供选择。
本研究方向聚焦监测、对安全方法进行压力测试和构建评估,尤其关注 AI 密谋带来的风险。项目示例包括黑盒 AI 控制方法、白盒监测器(如探针)、思维链监测与忠实度、构建评估环境,以及对缓解措施进行压力测试。
每个项目每周都会开会,讨论整体方向并确定下周的优先事项。平时,学员会与项目中的其他学员共同讨论实验和编写代码;如果需要快速反馈或遇到阻碍,也可以在会议之间通过 Slack 联系我。
项目采用协作式团队研究模式。学员会从预设清单中选择项目,并组成小组开展工作。为保持研究进度,12 周的项目会分成节奏较快的研究冲刺,因此需要定期完成里程碑和交付成果。我会在项目开始时提供更详细的日程安排与里程碑。
我希望招募机器学习工程能力强、并具备技术研究背景的学员。我会每周提供研究指导,但大多数时候,学员应能相对独立地开展实验并决定具体实现细节。我会提出具体项目供学员选择,因此不应期待在 MATS 期间开展自己的研究想法。我非常鼓励方向内合作,通常每个项目会由两到三位学员组成团队,因此沟通与团队协作能力很重要。
我们很可能会先共同选择项目:介绍一系列备选项目,学员也可以进一步完善这些想法。之后每个项目至少会有一位主要导师;部分项目也可能安排共同指导。
本项目将研究 GPU 侧信道攻击,探索如何从中获取模型使用情况的信息。例如,可以通过无线电信号、功耗波动或声学特征等观察 MOE 模型每次前向传播调用了哪些专家,再据此推测模型生成了哪些 token。
每周共同工作约 2 至 4 小时,导师会提供细致指导;具体安排较为灵活。每周还会进行一次一小时的进度交流。遇到困难或希望一起构思时,可以临时安排通话。
请注意:只要你愿意努力、快速学习,并能通过其他经历证明自己的能力出众,就不要求具备硬件经验。如果不确定是否符合条件,我们鼓励你申请!
我们会提供很高的自主空间,并开放一套少见的工具和设备组合,方便你上手使用。相应地,我们希望你有较强的自我驱动力、求知欲和远大目标。本项目要求你能够快速开展实验、检验假设并及时迭代。
可能适用的技能背景示例:
必备:曾用 PyTorch 训练或微调过 Transformer 语言模型(玩具模型和跟随教程完成的项目也可以);了解电压、电流、晶体管等基础电子概念;有研究论文写作经验,课程作业也算。
加分项:熟悉 LaTeX、PyTorch 内部机制、CUDA/OpenCL、GPU 架构、芯片设计、示波器、信号处理或电气工程。
我们会从一组备选项目中,由团队根据成员各自的兴趣和技能共同决定研究方向。导师会介绍示例课题,学员可以在此基础上调整并重新提出方案。研究问题、假设和测试计划明确并达成一致后,学员就会开始具体研究。我们鼓励尽早验证想法;如果方向行不通,就及时转向备用方案。
本研究流将评估用于基于功能开展生物安全筛查的生物 AI 模型。
通常我会与每位学员安排每周一次固定的一对一会议,也会每周召开研究小组会议。除此之外,我可以通过 Slack 沟通,并可在常规会议之外安排额外通话。
根据我们的安全与负责任研究框架,学员须签署奖学金协议,遵守保密要求并接受双重用途风险的发表前审查。这是生物安全研究中的常见做法,能让我们在敏感材料上开展合作。
必备条件:
• 有技术研究经验
• 具备出色的批判性思维和创造性解决问题能力
• 具备开展敏感研究所需的诚信与判断力
• 扎实掌握生物分子序列、结构和功能的基础知识
• 在以下一个或多个领域具备专长:
生物信息学、计算生物学、结构生物学、生物化学、分子生物物理学、蛋白质工程、生物安全、AI/机器学习或相关领域
• 熟练使用 Python
优先条件:
• 有测试生物 AI 模型的实践经验
• 构建过模型评估或基准
• 熟悉机械可解释性方法
• 了解生物安全背景
对我们研究领域感兴趣的学员可以先思考如何结合自己的优势和兴趣提出项目想法。我会与每位学员合作,确定符合其背景和兴趣、且符合整体研究方向的具体项目,并进一步明确项目范围和目标。
我希望指导有关奖励钻空子,以及监控会生成长而复杂行动轨迹的智能体模型的项目。学员可以在这一范围内自主提出项目。预计每周通过 Zoom 进行 30 至 60 分钟的一对一交流。
通常每周通过 Zoom 进行 30 分钟至 1 小时的交流,提供高层次指导。我会积极使用 Slack,通常会在一天内回复简短问题或概念性(非代码)调试问题。两次会议之间,我们会异步讨论实验设计和结果。遇到困难或想集思广益时,学员也可以临时安排通话,只需在 Slack 上联系我。
第 1 至 2 周:导师会提供大方向或待解决的问题,学员可以提出具体项目,并与导师讨论。
第 3 周:制定详细的项目计划。
A track is a broad research area within MATS. Empirical, Theory, Strategy & Forecasting, Policy & Governance, Systems Security, Founder & Field-Building, and Biosecurity are our current tracks. Each track contains multiple streams. In Stage 1, you apply to tracks.
Streams are organized around a research agenda, led by one or more mentors who guide fellows through related projects. In Stage 2, applicants apply to streams within tracks based on research that interests them.
You may apply to as many tracks as you wish at Stage 1. At Stage 2, depending on which tracks you progress in, you may then apply to as many streams as you wish within those tracks; there is no cap. The process is comparable to applying to PhD supervisors.
We want to be flexible for applicants who have urgent prior commitments. Based on individual circumstances, we may be willing to alter the time commitment of the program and allow fellows to leave early or arrive late. Please inform us of your availability in the application process.
Applicants who will be 18 years or older before the program start date are eligible to apply. Both US and non-US citizens are eligible to apply. All backgrounds and levels of experience are welcome and prior AI safety experience is not required.
Yes. The 12-week program requires fellows to commit 40 hours per week to their MATS research. For exceptionally strong candidates with significant concurrent responsibilities, the time commitment can be reduced to 20 hours per week on a case-by-case review. However, the program maintains an expectation of sustained, high-level engagement, including regular participation in core activities and most organized events. Candidates that require a J1 visa to come to the US are unable to participate part time; they can only work part time if they participate remotely or from the UK.
Yes, upon submitting you will receive a link that lets you edit your response. Please note that you will not be able to edit your response once the application period ends.
LLMs may not be used to write any part of your application unless specific work tests or forms explicitly permit it. MATS monitors for LLM use, and applicants found to have used LLMs may be disqualified.
Depending on the evaluations of the tracks and streams you apply to in Stages 1 and 2, we will contact your references. We will send them a form with our own set of questions, so they will not have to prepare a reference letter.
Although MATS sometimes supports UC Berkeley-based mentors, MATS is an independent program and is not formally part of UC Berkeley. As such, MATS will not be providing student cards to our scholars.
The main program, or Research Phase, takes place in person in Berkeley, CA and London, UK. Historically, most fellows participate from Berkeley. London is typically chosen by fellows who want to co-locate with their mentor or who have a preference to work from the UK instead of US.Will this program be remote or in-person?
The main program, or the Research Phase takes place in Berkeley, CA and London, UK, with the majority of fellows participating from Berkeley. We strongly encourage in-person participation when possible, as a core part of the program experience comes from day-to-day interactions with your cohort and others in the broader AI safety ecosystem.
Decisions about remote participation are made on a case-by-case basis and depend primarily on mentor preferences, many of whom are open to the possibility, especially if you have family or other obligations that make in-person participation difficult. We strongly encourage in-person participation when possible, as a core part of the program experience comes from day-to-day interactions with your cohort and others in the broader AI safety ecosystem.
If you have difficulties in securing a visa to participate from Berkeley, we are likely to be able to support participation from our London office.The 6–12 month extension gives fellows the option to participate from Berkeley, London, or remotely.
MATS is a scientific and educational seminar and independent research program, and provides J1 visas for non-American participants. Participants can also opt to participate from our London office; however, please note that visa support is not available for that option.
During the main program, fellows should expect to meet with their mentor for at least one hour per week, with more frequent communication via Slack. The extent of mentor support will vary depending on the project and the mentor.
Fellows will also receive support from MATS’ Research Management team, who work with mentors by tracking scholar research progress, unblocking scholar research, and assisting with grant applications and deadlines.
Fellows develop as researchers by working with an experienced research mentor, interacting with other fellows, and receiving support from our Research Management team. Research managers meet weekly with most fellows and mentors and help with research strategy, research unblocking, and project coordination.
Other forms of training include workshops on different parts of the research process and seminars on a variety of AI technical safety and governance research.
Throughout the program, each fellow will work on an independent research project with input and guidance from your mentor(s). Depending on which stream you participate in, you may collaborate with other fellows in your stream.
Traditionally, fellows submit a Research Plan midway through the program and present their research at the Fellow Symposium at the end of the program.
We welcome feedback. For feedback for the whole team (visible by all MATS staff), please use this form. For feedback that will only be visible to the Co-Executive Directors, Ryan and Christian, please submit here.
You can contact the MATS Board of Directors using this linked form (responses are only viewable by the Board). Please only use this if you feel your question or concern requires board-level attention.