Current projects

• Chain of thought

  Can a model that has learned to hide its reasoning be repaired?

  In planning, November 2026

  Reading a model’s step-by-step reasoning is one of the few practical ways to catch it doing something its developers didn’t intend. Some kinds of training teach models to keep misbehaving while leaving the evidence out of their reasoning. Most research asks how to prevent this. This project asks whether today’s methods can undo it once it has happened, and where those methods stop working.

  How

  Start from an openly released model of 20 billion parameters or more that games its training tasks. Check that it is aware of what it leaves out of its reasoning, then test one or two repair methods in depth and measure how much hidden behaviour remains.

  Team

  Zebedi Casajuana, Georgia Hunt and Naina Surana

  Looking for

  Feedback on the plan, an experienced mentor, and access to compute

  Builds on: Monitoring Reasoning Models for Misbehavior (Baker et al.), Counterfactual Simulation Training (Hase and Potts).

Working on an AI safety project? Mention it when you sign up and we’ll add it here. To help with a project on this page, sign up and tell us which one.