We introduce TRoVe, an automated approach for discovering error-inducing static feature biases learned by temporal VLMs.