Request for Proposals, AI Interpretability (2026)

• Receive funding to develop interpretability methods that detect deceptive behaviours in LLMs and steer their reasoning to eliminate these behaviours.
• Develop tools for detecting deceptive behaviours where model outputs contradict internal representations.
• Create steering methods for intervening on model truthfulness using mechanistic understanding.
• Apply detection and steering techniques to real-world use cases and human-AI teams.
• Evaluate methods on realistic scenarios beyond academic benchmarks to prove generalisation.

• Receive funding to develop interpretability methods that detect deceptive behaviours in LLMs and steer their reasoning to eliminate these behaviours.
• Develop tools for detecting deceptive behaviours where model outputs contradict internal representations.
• Create steering methods for intervening on model truthfulness using mechanistic understanding.
• Apply detection and steering techniques to real-world use cases and human-AI teams.
• Evaluate methods on realistic scenarios beyond academic benchmarks to prove generalisation.

Schmidt Sciences is a philanthropic organisation dedicated to fostering the advancement of science and technology. Its focus areas include AI and Advanced Computing, Astrophysics and Space, Biosciences, Climate, and Science Systems programs.