XIAOMI(HK:01810) Sou Hu Cai Jing·2026-02-03 03:35Core Insights - Xiaomi's founder and CEO Lei Jun announced that multiple research achievements from the Xiaomi team have been selected for ICLR 2026, focusing on areas such as multimodal reasoning, reinforcement learning, GUI agents, end-to-end autonomous driving, and audio generation [1] Group 1: Reinforcement Learning - The Xiaomi team's research titled "Shuffle-R1" introduces a dynamic data reorganization framework that addresses challenges in multimodal large model training, significantly improving gradient signal quality while surpassing existing reinforcement learning baselines with minimal computational overhead [2] Group 2: Mobile Intelligent Agents - The "MobileIPL" framework developed by the Xiaomi team pioneers iterative preference learning, optimizing thinking steps at a granular level and overcoming the scarcity of high-quality trajectories, achieving record performance in mainstream GUI-Agent tests [4] Group 3: End-to-End Autonomous Driving - The "ReCogDrive" research integrates innovative technologies by injecting prior driving knowledge into a hierarchical cognitive data pipeline, utilizing a cognitive-guided diffusion planner to generate physically feasible trajectories, and introducing the DiffGRPO reinforcement learning algorithm to directly optimize driving strategies, leading in closed-loop tests [5] Group 4: Other Innovations - Additional innovations from the Xiaomi team include "ThinkOmni," which enables zero-cost transfer of text reasoning capabilities to all modalities; "Flow2GAN," which combines flow matching and adversarial generation for high-fidelity audio synthesis; and "WorldSplat," which advances 4D driving scene generation technology [5]