This work investigates the effectiveness of reinforcement learning (RL) for small-scale models in graphical user interface (GUI) automation tasks. We systematically study the impact of different reward functions, as well as the trade-off between supervised fine-tuning (SFT) and RL data allocation. Furthermore, we demonstrate that strong GUI task performance can be achieved without relying on Chain-of-Thought (CoT) reasoning. Experimental results show that, on Odyssey dataset, a 2B-parameter model—without any access to historical states—achieves performance comparable to, and in some metrics surpassing, that of a 7B model. On Aitz dataset, our method improves the overall exact accuracy by 13% over previous models that leverage Chain-of-Action-Thought (CoAT) semantic reasoning. These findings highlight the potential of lightweight, context-free RL-based GUI agents for efficient and scalable deployment.