August 29, 2026 – Xiaomi has successfully secured the acceptance of eight research papers at EMNLP 2026, one of the most prestigious international conferences in the field of Natural Language Processing (NLP). The conference is scheduled to take place in Budapest, Hungary, from October 24 to 29. Many of these research achievements were accomplished in collaboration with multiple domestic universities, and the resulting technologies are expected to be integrated into Xiaomi’s “Human-Car-Home” smart ecosystem in the future.
The accepted papers cover several cutting-edge areas, including mobile AI agents, large language model (LLM) inference acceleration, AI-assisted R&D, and proactive intelligence. A significant portion of the research focuses on optimizing smartphone GUI (Graphical User Interface) agents. Traditionally, AI assistants have struggled with errors during automated app operations and required extensive manual rule-writing. Xiaomi’s new solution automates task generation and operation validation, eliminating the need for manual logic and achieving an accuracy rate of over 90%. Additionally, a novel two-step framework utilizes a lightweight model to determine whether to proactively prompt the user, preventing disruptive pop-ups and reducing computational overhead by 69.3%.
Another cluster of breakthroughs targets the acceleration of large models and computational efficiency. Xiaomi has developed a speculative decoding method that boosts inference speed by 2.4 to 3.01 times without adding any new parameters. Furthermore, an intelligent resource allocation mechanism has been introduced to optimize “deep thinking” processes; it dynamically assigns more computing power only to complex problems, significantly saving token consumption with minimal impact on performance. In the realm of AI-assisted R&D, a new framework can proactively filter out low-value model modifications, saving an average of 70.88% in GPU computing time and shortening the development cycle for edge AI models.
Unlike purely theoretical academic work, these research initiatives are specifically aimed at solving practical pain points in real-world scenarios, such as local deployment constraints and unreliable AI assistant operations. While the acceptance of these papers marks a crucial pre-accumulation of technology, Xiaomi acknowledges that transitioning from academic research to tangible consumer features requires further engineering refinement. These innovations will be progressively integrated into Xiaomi devices to continuously enhance the user experience of its AI assistants.
