Qinghua Xu

DreamTest adapts a recurrent state-space model into a testing world model that learns agent behaviour and environment dynamics from a DRL agent's training log, then scores a candidate configuration by generating imagined rollouts in parallel and combining termination-weighted failure predictions, guiding search without executing the simulator; on Parking, Humanoid and DonkeyCar its mean AUPRC exceeds the strongest baseline by about 97%, 12% and 39%, gains on five out-of-distribution sets reach 145%, 29% and 44%, and under the same simulator-validation budget the best DreamTest-plus-search combinations find 29%, 22% and 79% more novel failures on average.

DreamTest adapts a recurrent state-space model into a testing world model that learns agent behaviour and environment dynamics from a DRL agent's training log, then scores a candidate configuration by generating imagined rollouts in parallel and combining termination-weighted failure predictions, guiding search without executing the simulator; on Parking, Humanoid and DonkeyCar its mean AUPRC exceeds the strongest baseline by about 97%, 12% and 39%, gains on five out-of-distribution sets reach 145%, 29% and 44%, and under the same simulator-validation budget the best DreamTest-plus-search combinations find 29%, 22% and 79% more novel failures on average.