Auditing language models for hidden objectives

Sam Marks, Johannes Treutlein, Trenton Bricken, Jack Lindsey, Evan Hubinger · 2025 · arXiv

Read the paper

Trains a model with a hidden objective and runs a blind auditing game: three of four teams recover the objective using interpretability, behavioral attacks and training-data analysis. (Author list abbreviated; full list at the link.)

Questions it bears on

• Do models represent goals and preferences
• An agenda beyond the Assistant