star2vec

When talked into harm, a model blames the an­swer, not it­self (an in­ter­pretabil­ity study of guilt vs shame)

No sign of back­track­ing in la­tent rea­son­ing: the fi­nal an­swer sim­ply set­tles in instead

The im­posters among us: func­tion vec­tors that ace ev­ery check and do the wrong task (in search of cir­cu­lar­ity)

Whack-a-mole with a bro­ken ham­mer: does a model in­ter­nally track its au­toma­ton state?