1 paper
David Vella Zarb, Rustem Turtayev, Taywon Min +2
Safety evaluations often infer latent motivations from behavioral patterns, but the construct validity of these inferences is unclear. We study this problem in alignment faking, wh…