Research
Alignment Research
Founded in early 2026, Arcadia Alignment is a London-based group of AI safety researchers. We study the science behind frontier approaches to AI alignment and seek to produce legible evidence for safe decision-making.
Our Research
AI is increasingly capable of AI research and development. We study where automated alignment research stands today, where it fails, and where it differs from capabilities.
The effects of large-scale reinforcement learning on model values are poorly understood. We investigate how training shapes model motivations and when it may erode alignment.
Featured work
Featured work
We stress-test alignment midtraining (AMT) across model and token budget scales. Our results suggest that midtraining cannot tackle the hard problems of AI alignment—namely distributional shift and reward underspecification in the presence of imperfect data.
Show more↓
Blog
Debate with Self-Play Best-of-N Optimization
Jul 9, 2026
Blog
Your Model Organisms Might Be Fried
Jun 18, 2026
Our Team
Opportunities
The alignment team is not currently hiring. View our other open roles, or sign up to our mailing list to hear when new positions open.
Mailing list
Stay updated on our work.
Questions?
Reach out via our contact form.