Ben studies evaluation awareness: when models recognize that they’re being tested. Current safety assessments rely heavily on behavioral evaluations, and if models behave differently when they know they’re being tested, these assessments could be systematically misleading. Ben’s goal is to determine the extent of this issue and its implications for AI safety.
Before FAR.AI, he was a postdoc at the Center for Human-Compatible AI at UC Berkeley, mentored by Stuart Russell. His research centered on generalization: how models handle unfamiliar inputs. He especially focused on designing agents that recognize when they’re in unfamiliar situations and ask for help. Before pivoting to AI safety, he worked on incorporating driver preferences into ride matching at Lyft. He holds a PhD in Computer Science from Stanford University, where he studied fairness in markets and algorithms, advised by Ashish Goel.
