CoME: Empowering Channel-of-Mobile-Experts with Informative Hybrid-Capabilities Reasoning

Mobile Agents requires hybrid-capabilities reasoning, including screen summary, subtask planning, action decision and action function. However, existing agents struggle to achieve both decoupled enhancement and balanced integration of these capabilities. To address these challenges, we propose Channel-of-Mobile-Experts (CoME), a novel agent architecture consisting of four distinct experts, each aligned with a specific reasoning stage, CoME activates the corresponding expert to generate output tokens in each reasoning stage via output-oriented activation. To empower CoME with hybrid-capabilities reasoning, we introduce a progressive training strategy: Expert-FT enables decoupling and enhancement of different experts’ capability; Router-FT aligns expert activation with the different reasoning stage; CoT-FT facilitates seamless collaboration and balanced optimization across multiple capabilities. To mitigate error propagation in hybrid-capabilities reasoning, we propose InfoGain-Driven DPO (Info-DPO), which uses information gain to evaluate the contribution of each intermediate step, thereby guiding CoME toward more informative reasoning. Comprehensive experiments show that CoME outperforms dense mobile agents and MoE methods on both AITZ and AMEX datasets.

Mobile agents are AI systems that operate smartphone apps by reading the screen and deciding what to do next. To complete a task, they need several abilities at the same time: understanding what is on the screen, planning the next step, choosing the right action, and calling the correct function. Existing agents often mix these abilities together, which can make some skills undertrained and cause mistakes to spread through the reasoning process. We propose Channel-of-Mobile-Experts (CoME), a mobile agent that assigns different expert components to different reasoning steps. This lets each expert focus on the part of the task it is best suited for, while still working together with the others. We also design a step-by-step training strategy to first strengthen each expert and then teach them to collaborate. Finally, we use information gain to identify which intermediate reasoning steps are most useful, so the model can learn from better reasoning paths.