[Part-time AI Safety Research Program] MARS 3.0 Applications Open for Participants & Recruiting Mentors

By Cambridge AI Safety Hub @ 2025-05-07T22:52 (+3)

The Cambridge AI Safety Hub (CAISH) is now accepting applications for MARS 3.0 (Mentorship for Alignment Research Students), a volunteer-based program connecting aspiring researchers with experienced mentors to conduct AI-safety research.

The program runs from July 7 to late September 2025, beginning with an intensive week in Cambridge, UK, followed by remote collaboration.

Previous participants have published at NeurIPS, ICML, EMNLP, and ControlConf, with mentorship from researchers at Redwood Research, Apollo Research, Google DeepMind, the University of Cambridge & other organisations.

Key Details:

• Focus Areas: Technical and policy domains in AI safety.
• Time Commitment: Approximately 8–15+ hours per week.
• In-Person Week: July 7–13 in Cambridge, with accommodation and travel support provided.
• Mentorship: Weekly meetings with mentors and research managers.
• Community / Network: Workshops and presentations from relevant speakers and organisations during the in-person week, and possibly afterwards.

Application Process:

• Deadline: May 18, 2025, 23:59 AoE
• Stages: Initial application → coding test or written assignment (for some participants) → mentor-specific assignments → decisions by June 16
• Eligibility: Open to students and professionals, especially those with backgrounds in computer science, mathematics, cybersecurity, hardware engineering, policy, political or social sciences.

Calling Prospective Mentors

Find out more at our page for prospective mentors.

• Who we’re looking for: 
  Researchers with well-scoped projects excited to mentor fresh talent in the field.
  In our experience dedication and a clear vision for the project from mentors is a key factor in the success of the projects. That said, we are excited to receive applications from up-and-coming mentors as well who have reason to believe could execute well on leading a project.

  Some areas we are especially excited about include: 
  (Meant to be interpreted as our thinking about currently most pressing areas, but we are broadly interested in mentors pursuing other research directions as well.)

  • Evaluations that fit well into policy frameworks (e.g. R&D evals, informing timelines models)
  • AI control
  • Model organisms / demos for misalignment & scheming
  • Chain-of-thought faithfulness & steganography
  • Hardware-enabled mechanisms
  • Model-weight security and cybersecurity broadly
  • AI Governance, broadly & especially moving towards establishing red lines, if-then commitments or aiding international cooperation (e.g. via credible commitments).
  • Formal verification applications
• What to expect as a mentor? 
  MARS is currently closest to in format and applicant profile to SPAR, with the main difference being the additional in-person component.  (Which has had broadly positive reviews both from mentors and participants in previous iterations in giving projects momentum.) 
  
  As a mentor, you will make the decision on which applicants you wish to work with . We encourage you to feel comfortable withdrawing from the program if you don't find great collaborators.
• Compensation: $30 USD/hour for your time (≈10 hours during the in-person week + weekly check-ins afterwards).
• Logistics: Kick-off July 7–13 in Cambridge (travel & accommodation covered); remote mentoring thereafter.  (Not a dealbreaker if you can't attend this as a mentor.)

For more information and to apply, visit:
http://caish.org/mars