Nathaniel Mitrani

What Drives the Com­pli­ance Gap? A Three-Driver De­com­po­si­tion of Align­ment Faking

Char­ac­ter-trained mod­els can strug­gle to generalise

Learned Chain-of-Thought Obfus­ca­tion Gen­er­al­ises to Unseen Tasks

In­ves­ti­gat­ing Neu­ral Scal­ing Laws Emerg­ing from Deep Data Structure

Mak­ing the case for av­er­age-case AI Control