aditichaudharyy14@gmail.com

Does context framing change how a model verbalizes a planted hint? Aditi Chaudhary. MATS 12.0 application, Neel Nanda stream *Code and data: https://github.com/AditiChaudharyy14/cot-hint-verbalization * Time: 17.3 hours, plus 1.2 hours Executive summary CoT is faithful when the answer came from the stated reasoning; unfaithful when the reasoning is a post-hoc...

What I built I built a small harness to test whether an AI agent's policy limits actually hold under attack. It runs on my laptop against a local model, so the experiment cost nothing. I count a finding when the agent attempts a tool call that policy should have blocked,...