MATS Summer 2026

In this stream we will explore extensions and implications of our discovery that neural networks pretrained on next-token prediction represent belief-state geometry in their activations. We will build on this fundamental theory of neural network representations in order to discover what AI systems are thinking, and understand their emergent behaviors.

Peter Henderson’s stream focuses on developing safe, aligned AI agents, with projects on scalable oversight rules informed by law and game theory, safe long-horizon exploration, and measuring “jagged” capability/safety frontiers. Scholars will join an independently driven, engineering-heavy research environment, collaborating with other MATS scholars and PhD students, with weekly 1:1s and active async mentorship.

The Redwood Research stream is looking for fast empirical iterators and strategists to work on control research.

My MATS fellows will do philosophical thinking about multi-agent intelligence and how agents change their values. This will likely involve trying to explore and synthesize ideas from game theory, signaling theory, reinforcement learning, and other related domains.

Roger Grosse’s stream investigates how to improve influence functions and other training data attribution methods, and uses these tools to study alignment-related phenomena such as out-of-context reasoning and emergent misalignment. The ideal scholar has experience with LLM internals, strong statistics/applied math skills (especially numerical linear algebra), and can independently drive research from literature review through experimentation and analysis. Roger provides shovel-ready projects while giving exceptional scholars freedom to pursue their own ideas, and is open to scholars collaborating with others.

International coordination to reduce frontier AI risks, with a focus on China and the West. 

We build scalable technology for AI understanding and oversight.

The stream will focus on conceptual, empirical, and theoretical work on scalable oversight and control. This includes but is not limited to creating model organisms for specific failure modes, designing training procedures against them, and making progress on subproblems involved in safety cases.