Imitation Learning has hit a structural ceiling. Researchers from Meituan and Fudan University have correctly identified a fatal flaw in the classic approach: static datasets cannot capture the causal feedback loop of a real-world operating system environment. In a live OS, every action radically alters the screen state and the range of available options. Models trained simply to "mimic" logs are helpless when faced with errors that accumulate over long-range, multi-step tasks. It is time to admit that the era of offline training on frozen trajectories is giving way to an era of active experience, where agents learn on the job.
The EvoCUA-1.5 framework is an attempt to move beyond simple language-based recipes and master complex tasks in partially observable multimodal environments. At the core of the system is Step-Level Policy Optimization (STEPO), a method that decomposes long action chains into individual samples while maintaining a balance of advantages. To stabilize the process, the authors introduced the Dynamic Tri-Adaptive Curriculum (DTAC). This curriculum juggles manageable tasks, complex replays of successful scenarios, and controlled exposure to impossible cases. The entire process relies on an asynchronous RL infrastructure designed specifically for slow environment response times and the inherent variability of OS workflows.
Key features of the EvoCUA-1.5 architecture:
The STEPO method for efficient decomposition of long action sequences. The DTAC three-tier curriculum, adapting complexity to the agent's current skill level. An asynchronous training environment optimized for the specific latencies of operating systems.
Technical results confirm that the bet on active interaction has paid off. EvoCUA-1.5 achieved a 63.2% score on the OSWorld-Verified benchmark. According to the report, this performance allows the model to outperform open-source counterparts in the 32B–35B range and even compete with heavyweights that significantly exceed it in parameter count. By focusing on verifiable results in executable sandboxes rather than blindly copying human actions, the framework grants agents the ability to self-correct.
The era of primitive macros and hard-coded scripts is ending. For business leaders and product owners, the signal is clear: value is shifting from accumulating massive datasets of human actions to creating robust simulation environments. The future of enterprise software belongs to autonomous systems capable of learning from their own failures in real time, rather than waiting for the next database update.