Mentorship for Alignment Research Students (MARS)
About
MARS (Mentorship for Alignment Research Students) is a research mentorship program run by the Cambridge AI Safety Hub (CAISH), itself a project of Meridian Impact CIC, a Community Interest Company registered in the UK. MARS was founded in 2024 by Chloe Li, then director of CAISH and now an Anthropic Fellow. The program is currently co-directed by Gaurav Yadav and Justin Dollman. The program pairs small teams of 2-4 early-career researchers or students with experienced mentors to produce published AI safety research. The structure combines an intensive in-person kick-off week in Cambridge with a subsequent 10-12 week remote research phase, preceded by 2-4 weeks of preparatory work. Participants commit approximately 8-15+ hours per week throughout. Mentors receive $30/hour for their time during the in-person week and ongoing check-ins. MARS operates across two main tracks: technical AI safety (covering AI control, mechanistic interpretability, evaluations, and chain-of-thought faithfulness) and AI governance and policy (covering international agreements, compute governance, and hardware-enabled mechanisms). The program is completely free to participants; CAISH provides each team with a $2,000+ compute budget, Claude Max access for technical streams, travel funding and accommodation for the sprint week, and dedicated research management support. As of early 2026, MARS has run five cohorts (MARS I through MARS V), with MARS VI planned for December 2026. Alumni have published research at NeurIPS 2025, ICML, EMNLP, and ControlConf, and have gone on to positions at Apollo Research, Anthropic, AISI, LASR Labs, MATS, and Geodesic Research. A notable output from MARS I is the paper 'Compact Proofs of Model Performance via Mechanistic Interpretability' (arXiv 2406.11779), published in June 2024. MARS III produced two papers including 'Depth-Wise Activation Steering for Honest Language Models' and 'Combining Cost-Constrained Runtime Monitors for AI Safety,' both presented at NeurIPS 2025. Across all cohorts, over 90 students and mentors have participated. A parallel program, MARS London, is run by Safe AI London in collaboration with CAISH. CAISH receives funding from Open Philanthropy for its operating costs and is hosted within the Meridian Cambridge hub at 53-54 Sidney Street, Cambridge.
Theory of Change
MARS aims to reduce existential risk from advanced AI by expanding the pipeline of capable AI safety researchers. By pairing early-career researchers with established mentors on well-scoped, publishable projects, MARS accelerates participants' transition into full-time AI safety work. The program bets that a major bottleneck is talent: there are many people with relevant technical or policy backgrounds who could contribute to AI safety but lack mentorship, research experience, and community. By providing structured mentorship, compute resources, and an in-person collaborative environment, MARS lowers the barriers to producing credible safety research and helps participants develop the skills and networks needed for longer-term impact. Alumni placements at organizations like Anthropic, AISI, Apollo Research, and MATS represent the downstream field-building effect.
Details
- Start Date
- -
- End Date
- -
- Expected Duration
- -
- Funding Raised to Date
- -
- Last Updated
- Apr 3, 2026, 1:20 AM UTC
- Created
- Apr 3, 2026, 1:20 AM UTC