EvoCUA-1.5: Online Reinforcement Learning for Multi-turn Computer-Use Agents

EvoCUA-1.5 advances computer-use agents by transitioning from offline experience learning to online reinforcement learning within executable sandbox environments. This approach addresses the causal feedback loops of multimodal desktop environments where actions alter screen states. Similarly, the offline-to-online reinforcement learning (O2O-RL) paradigm allows policies trained on large datasets to be refined through limited online interaction. This is critical for nonstationary domains, although fine-tuning performance remains highly sensitive to specific algorithms and hyperparameters during deployment.