Ilya Nachevsky

Steganog­ra­phy via in­ter­nal ac­ti­va­tions is already pos­si­ble in small lan­guage mod­els — a po­ten­tial first step to­ward per­sis­tent hid­den rea­son­ing.

Sleep peace­fully: no hid­den rea­son­ing de­tected in LLMs. Well, at least in small ones.