This paper addresses the challenge of network resource inefficiency in traditional broadcast-based cooperative perception (CP) for connected and autonomous vehicles (CAVs). To improve efficiency, the authors propose an adaptive CP scheme for CAV pairs in a mixed-traffic autonomous driving scenario. The scheme dynamically selects CAV pairs to switch between standalone perception (SP) and CP modes based on dynamic perception workloads and channel conditions. The authors formulate a joint adaptive CAV cooperation and resource allocation problem, which is decoupled into an adaptive CAV cooperation subproblem and a series of instantaneous resource allocation subproblems. A model-assisted multi-agent reinforcement learning (MARL) solution is developed, integrating MARL for adaptive CAV cooperation decisions and a model-based solution for resource allocation. The proposed scheme aims to maximize computing efficiency gain while minimizing switching costs, ensuring that perception tasks meet delay requirements. Simulation results demonstrate the effectiveness of the proposed scheme compared to benchmark schemes.This paper addresses the challenge of network resource inefficiency in traditional broadcast-based cooperative perception (CP) for connected and autonomous vehicles (CAVs). To improve efficiency, the authors propose an adaptive CP scheme for CAV pairs in a mixed-traffic autonomous driving scenario. The scheme dynamically selects CAV pairs to switch between standalone perception (SP) and CP modes based on dynamic perception workloads and channel conditions. The authors formulate a joint adaptive CAV cooperation and resource allocation problem, which is decoupled into an adaptive CAV cooperation subproblem and a series of instantaneous resource allocation subproblems. A model-assisted multi-agent reinforcement learning (MARL) solution is developed, integrating MARL for adaptive CAV cooperation decisions and a model-based solution for resource allocation. The proposed scheme aims to maximize computing efficiency gain while minimizing switching costs, ensuring that perception tasks meet delay requirements. Simulation results demonstrate the effectiveness of the proposed scheme compared to benchmark schemes.