AdaptArena finds web agents miss 67.3 points

AdaptArena found web agents reached at most 15.62% success across 480 tasks requiring them to infer users’ unstated preferences, versus 82.92% for oracle agents given the correct preference. The benchmark paper says downstream execution still failed even when an agent inferred the preference correctly, separating personalization from reliable action grounding. Teams shipping browser agents should add hidden-preference and recovery cases beside the archive’s warning about versioned computer-use scores before treating a clean navigation demo as user alignment.