Xiaomi lands 8 papers at EMNLP 2026, betting heavy on agents that use your phone
Xiaomi's accepted papers stretch from mobile agents to efficient training and inference. GSAR, GUI-PRA and MaP lay the groundwork for AI that operates the phone for you.
Xiaomi announced that 8 of its papers were accepted at EMNLP 2026, one of the most respected venues in natural language processing. The titles span four areas: mobile agents, efficient LLM training and inference, AI-assisted model development, and proactive intelligence. For a phone maker the message is clear: the company is doing the foundational research for AI that will operate the device on your behalf.
The mobile-agent trio is the most concrete part of the package. GSAR scales reinforcement learning for mobile agents with self-evolving tasks and goal-state anchor rewards; GUI-PRA acts as an active judge that verifies UI changes at every step; and MaP unifies action prediction, grounding and planning under a single masked-trajectory objective. Together they systematize both the training and the supervision of agents that look at a screen and tap it.
Context matters too: Chinese manufacturers are betting that the phone's next layer will be the agent. Xiaomi showing up at this intensity on the academic front reads as a preview of agent features heading to HyperOS; the company announced the batch on X with paper summaries.
