Empirical
Streams in this track include hands-on research using machine learning experiments to understand and improve model safety including AI control, interpretability, scalable oversight, evaluations, red-teaming, and robustness. This is the largest track in the program and is defined by its methods rather than any single research agenda. If your primary tool is ML engineering, this is your track.
- Initial application: No track-specific questions.
- Stage 2: Complete 1–2 assessments evaluating research taste and technical implementation skills.
- Stream applications & follow-up: Apply to individual streams; follow-up includes interviews or additional assessments depending on the stream.
The track is defined by its methodology more than by any single research agenda. Fellows run ML experiments to understand and improve the safety properties of frontier models, with work spanning interpretability, AI control, scalable oversight, evaluations, red-teaming, robustness, and model organisms of misalignment. The unifying thread is that progress comes from getting hands on real models (training, probing, fine-tuning, measuring) rather than reasoning from first principles alone. This is the largest track in the program and the most common entry point into technical AI safety research.
We are looking for fellows whose primary tool is ML engineering, broadly construed. The essential requirement is the ability to design and run experiments on language models or other deep learning systems and iterate quickly on the results. In practice that usually means strong Python (with and without AI coding tools), comfort with the infrastructure around running models at moderate scale, and enough research taste to know which experiments are worth running. Mission alignment matters: fellows should be able to say why a given line of empirical work meaningfully reduces frontier risk, not just whether it yields a successful publication. Educational background and seniority are weighted lightly here relative to other tracks. Past cohorts have included strong fellows ranging from undergraduates to senior industry researchers.
Fellows are matched to mentors based on fit, and projects are scoped to produce concrete artifacts by program end: papers, evaluation suites, open-source tooling, or technical reports. Target audiences include safety and alignment teams at frontier labs, governments and other evaluation organizations, the broader ML research community.
Neel takes a pragmatic approach to interpretability: identify what stands between where we are now and where we want to be by AGI, and then focus on the subset of resulting research problems that can be tractably studied on today's models. This can look like diving deep into the internals of the model, or simpler black box methods like reading and carefully intervening on the chain of thought - whatever is the right tool for the job. This could look like studying how to detect deception, understanding why a model took a seemingly concerning action, or fixing weak points in other areas of safety, e.g. using interpretability to stop models realising they are being tested. You can learn more about Neel's approach in this podcast.
He has spent far too much time having MATS scholars, and has worked with ~60 so far - he’s excited to take on even more!
This stream focuses on empirical AI control research, including defending against AI-driven data poisoning, evaluating and attacking chain-of-thought monitorability, and related monitoring/red-teaming projects. It is well-suited to applicants already interested in AI safety with solid Python skills, and ideally prior research or familiarity with control literature/tools (e.g. Inspect/ControlArena).
Building realistic defensive cybersecurity benchmarks. Asymmetric Security responds to real cyber incidents and therefore holds data not available in the public domain. We would like to work with MATS scholars to build realistic benchmarks grounded in these real cyber incidents.
This coalition of mentors make up the “Anthropic Stream”. This stream spans a range of empirical research areas in AI safety on LLMs, including AI control, scalable oversight, model organisms, model internals, model welfare, security, and more. You’ll be pitched, and have the option to pitch, a variety of safety research projects, and then be matched to projects and mentors based on your interests/preferences on research and what you’d like to get out of MATS. Fellows in this stream frequently receive funding and continued mentorship after MATS to complete their research project, usually leading to a (co-)first author paper. People in this stream often end up in long-term homes for safety research after MATS (e.g. Anthropic, Redwood Research, OpenAI).
Anthropic mentors share an application, tend to collaborate and co-mentor projects together, and generally share infrastructure to streamline the fellow experience. By applying to this stream, you are being considered for all of the Anthropic mentors.
This coalition of mentors make up the “megastream”. This stream spans a range of empirical research areas in AI safety on LLMs, including AI control, scalable oversight, model organisms, model internals, model welfare, security, and more. You’ll be pitched, and have the option to pitch, a variety of safety research projects, and then be matched to projects and mentors based on your interests/preferences on research and what you’d like to get out of MATS. Scholars in this stream frequently receive funding and continued mentorship after MATS to complete their research project, usually leading to a (co-)first author paper. People in this stream often end up in long-term homes for safety research after MATS (e.g. Anthropic, Redwood Research, OpenAI).
Megastream mentors share an application, tend to collaborate and co-mentor projects together, and generally share infrastructure to streamline the scholar experience. By applying to this stream, you are being considered for all of the megastream mentors. In the application process, you can indicate particular mentors you are interested in working with.
Arthur Conmy's MATS Stream focuses on evaluating interpretability techniques on current and future AI Safety problems.
This can involve creating new safety techniques, as well as creating benchmarks and measuring performance against baseline techniques.
I have two broad areas.
Security:
I am interested in building demonstrations for hacking real-world AI deployments to show that they are not secure. The goal is to force companies to invest in alignment techniques that can solve the underlying security issues.
Benchmarks:
I am interested in building benchmarks to determine how generalizable modern LLM techniques actually are, now that we are no longer in the pre-training scaling era.