Martín Soto

I’m a research scientist on AI Alignment
at the UK government’s AI Security Institute.

This means I develop new research agendas to mitigate catastrophic risks from future AI. My current focus is automated Alignment.

Before that, I worked with Owain Evans on AI self-awareness,
did AI risk modelling with the Center on Long-term Risk,
and published research in Mathematical Logic and Decision Theory.

Email | CV | LinkedIn | Alignment Forum | Send me feedback

Publications in AI

• Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs. Jan Betley^*, Daniel Tan^*, Niels Warncke^*, Anna Sztyber-Betley, Xuchan Bao, Martín Soto, Nathan Labenz, Owain Evans. Oral at ICML 2025. [Website, Paper, Tweets]
• Tell me about yourself: LLMs are aware of their learned behaviors. Jan Betley*, Xuchan Bao*, Martín Soto*, Anna Sztyber-Betley, James Chua, Owain Evans. Spotlight at ICLR 2025. [Paper, Blog, Tweets]
• Backfire risks from advances in commitment technology. Duncan McClements, Urvi Gaur, Martín Soto^†.
• The marginal impact of AGI interventions. Martín Soto, Tristan Cook. Technical report.

* Equal contribution, ^† Mentorship role

Publications in Math

• Type-2 Feedback Computability. Juan Aguilera, Martín Soto. Forthcoming in Computability.
• Interactive proofs in Bounded Arithmetics. Martín Soto. Master’s Thesis, presentation.
• The real numbers in inner models of set theory. Martín Soto. Bachelor’s Thesis.
• Logically Updateless Decision-Making. Martín Soto, Abram Demski. Technical report, PIBBSS Symposium 2023.

Talks

• Evaluating AI Control mitigations, at Singapore Alignment Workshop. Recording.
• How can we effectively govern AI?, at Adevinta Barcelona. Post.
• Open-source Game Theory, at the Barcelona Logic Seminar. Recording.
• Logic & Music, at the Polytechnic University of Barcelona. Recording.