IT之家 2 月 3 日消息，小米创办人、董事长兼 CEO 雷军今日宣布，小米团队的多篇最新研究成果，成功入选 ICLR 2026，研究方向涵盖多模态推理、强化 学习、GUI Agent、端到端自动驾驶以及音频生成等领域。 IT之家注：ICLR（国际学习表征会议，全称是 International Conference on Learning Representations）是人工智能领域国际顶级会议之一，由图灵奖得主 Yoshua Bengio 和 Yann LeCun 于 2013 年创立的深度学习领域学术会议，致力推动人工智能理论与方法的前沿研究与创新发展。 小米本次入选国际顶级会议 ICLR 2026 的研究成果如下： 《Shuffle-R1: Efficient RL framework for Multimodal Large Language Models via Data-centric Dynamic Shuffle》 强化学习已成为提升多模态语言模型推理能力的重要后训练范式。然而，现有的强化学习训练流程在训练中仍面临效率低下的问题，其根源在于 两个长期被忽视的关键现象：优势坍缩（Advan ...

Core Insights - The article discusses advancements in Vision-Language-Action (VLA) models for autonomous driving, highlighting a shift from traditional supervised learning methods to reinforcement learning (RL) approaches to enhance model generalization and reasoning capabilities [2]. Summary by Sections VLA + RL Research Overview - The article summarizes recent works in the VLA + RL domain, indicating a trend towards using RL to address limitations in previous models, particularly in terms of hallucination issues and the efficiency of continuous action space exploration [2]. Key Papers and Contributions - **MindDrive**: Introduces a framework that transforms action space into a discrete language decision space, achieving a driving score of 78.04 and a success rate of 55.09% on the Bench2Drive benchmark using a lightweight model [6]. - **WAM-Diff**: Proposes an end-to-end VLA framework that utilizes masked diffusion for trajectory optimization, achieving superior performance on the NAVSIM benchmark [7]. - **LCDrive**: Addresses temporal expression and latency issues in text chain reasoning by employing a latent chain-of-thought mechanism, demonstrating improved reasoning efficiency and trajectory quality [12]. - **Reasoning-VLA**: Develops a framework that enhances parallel trajectory generation through learnable action queries, achieving high performance across multiple datasets [13]. - **Alpamayo-R1**: Bridges reasoning and action prediction through a modular architecture and multi-stage training, improving generalization in long-tail scenarios [18]. - **AdaThinkDrive**: Introduces a dual-mode mechanism to balance decision accuracy and reasoning efficiency, achieving a PDMS score of 90.3 on the Navsim benchmark [20]. - **AutoDrive-R²**: Combines supervised fine-tuning and RL to enhance trajectory planning accuracy, achieving state-of-the-art performance with a significant reduction in error rates [25]. - **IRL-VLA**: Proposes a framework that avoids reliance on simulators by using a reward world model, achieving state-of-the-art performance on the NAVSIM v2 benchmark [31]. - **DriveAgent-R1**: Integrates active perception with hybrid thinking, achieving significant improvements in decision reliability and efficiency [32]. - **Drive-R1**: Connects reasoning and planning in VLMs, providing effective methods for integrating reasoning with motion planning [37]. - **ReCogDrive**: Merges cognitive reasoning with diffusion planners, achieving state-of-the-art performance while addressing the limitations of imitation learning [38].

Core Insights - The focus of academia and industry has shifted towards VLA (Vision-Language Action) in autonomous driving, which provides human-like reasoning capabilities for vehicle decision-making [1][4] - Traditional methods in perception and lane detection have matured, leading to decreased attention in these areas, while VLA is now a critical area for development among major autonomous driving companies [4][6] Summary by Sections Introduction to VLA - VLA is categorized into modular VLA, integrated VLA, and reasoning-enhanced VLA, which are essential for improving the reliability and safety of autonomous driving [1][4] Course Overview - A comprehensive course on autonomous driving VLA has been designed, covering foundational principles to practical applications, including cutting-edge algorithms like CoT, MoE, RAG, and reinforcement learning [6][12] Course Structure - The course consists of six chapters, starting with an introduction to VLA algorithms, followed by foundational algorithms, VLM as an interpreter, modular and integrated VLA, reasoning-enhanced VLA, and a final project [12][20] Chapter Highlights - Chapter 1 provides an overview of VLA algorithms and their development history, along with benchmarks and evaluation metrics [13] - Chapter 2 focuses on the foundational knowledge of Vision, Language, and Action modules, including the deployment of large models [14] - Chapter 3 discusses VLM's role as an interpreter in autonomous driving, covering classic and recent algorithms [15] - Chapter 4 delves into modular and integrated VLA, emphasizing the evolution of language models in planning and control [16] - Chapter 5 explores reasoning-enhanced VLA, introducing new modules for decision-making and action generation [17][19] Learning Outcomes - The course aims to deepen understanding of VLA's current advancements, core algorithms, and applications in projects, benefiting participants in internships and job placements [24]

Core Viewpoint - The focus of academia and industry after end-to-end systems is on VLA (Vision-Language-Action), which provides human-like reasoning capabilities for safer and more reliable autonomous driving [1][4]. Summary by Sections Introduction to Autonomous Driving VLA - VLA is categorized into modular VLA, integrated VLA, and reasoning-enhanced VLA, which are essential for advancing autonomous driving technology [1][4]. Technical Maturity and Employment Demand - The demand for autonomous driving VLA solutions is high among major companies, prompting them to invest in self-research and development [4]. Course Overview - A comprehensive learning roadmap for autonomous driving VLA has been designed, covering principles to practical applications [4][6]. Core Content of Autonomous Driving VLA - Key topics include visual perception, large language models, action modeling, model deployment, and dataset creation, with cutting-edge algorithms like CoT, MoE, RAG, and reinforcement learning [6]. Course Collaboration - The course is developed in collaboration with Tsinghua University's research team, featuring detailed explanations of algorithms and practical assignments [6]. Course Structure - The course consists of six chapters, each focusing on different aspects of VLA, including algorithm introduction, foundational algorithms, VLM as an interpreter, modular and integrated VLA, reasoning-enhanced VLA, and a final project [12][20]. Chapter Details - Chapter 1 covers the concept and history of VLA algorithms, including benchmarks and evaluation metrics [13]. - Chapter 2 focuses on foundational algorithms related to Vision, Language, and Action, along with model deployment [14]. - Chapter 3 discusses VLM's role as an interpreter in autonomous driving, highlighting key algorithms [15]. - Chapter 4 delves into modular and integrated VLA, emphasizing the evolution of language models in planning [16]. - Chapter 5 explores reasoning-enhanced VLA, introducing new modules for decision-making and action output [17]. - Chapter 6 involves a hands-on project where participants build and fine-tune their models [20]. Learning Outcomes - The course aims to deepen understanding of VLA's current advancements and core algorithms, equipping participants with practical skills for future research and applications in the autonomous driving sector [22][26]. Course Schedule - The course is set to begin on October 20, with a structured timeline for each chapter's release [23]. Prerequisites - Participants are expected to have a foundational knowledge of autonomous driving, large models, reinforcement learning, and programming skills in Python and PyTorch [26].

Core Insights - The article discusses the advancements in Vision-Language-Action (VLA) models for autonomous driving, highlighting the integration of navigation and reinforcement learning to enhance reasoning capabilities beyond visual range [2][3][6]. Group 1: NavigScene - NavigScene is introduced as a novel auxiliary dataset that pairs local multi-view sensor inputs with global natural language navigation guidance, addressing the critical gap between local perception and global navigation context in autonomous driving [6]. - Three complementary paradigms are implemented in NavigScene: navigation-guided reasoning, navigation-guided preference optimization, and navigation-guided VLA models, enhancing the reasoning and generalization capabilities of autonomous driving systems [6]. - Comprehensive experiments demonstrate significant performance improvements in perception, prediction, and planning tasks by integrating global navigation knowledge into autonomous driving systems [6]. Group 2: AutoVLA - AutoVLA is proposed as an end-to-end autonomous driving framework that integrates physical action tokens with a pre-trained VLM backbone, enabling direct policy learning and semantic reasoning from raw visual observations and language instructions [12]. - A reinforcement learning-based post-training method using Group Relative Policy Optimization (GRPO) is introduced to achieve adaptive reasoning and further enhance model performance in end-to-end driving tasks [12]. - AutoVLA achieves competitive performance across multiple autonomous driving benchmarks, including open-loop and closed-loop tests [12]. Group 3: ReCogDrive - ReCogDrive is presented as an end-to-end autonomous driving system that integrates VLM with a diffusion planner, employing a three-stage training paradigm to address performance drops in rare and long-tail scenarios [13][16]. - The first stage involves fine-tuning the VLM on a large-scale driving Q&A dataset to mitigate domain gaps between general content and real-world driving scenarios [16]. - The method achieves a state-of-the-art PDMS score of 89.6 on the NAVSIM benchmark, highlighting its effectiveness and feasibility [16]. Group 4: Impromptu VLA - Impromptu VLA introduces a large-scale, richly annotated dataset aimed at addressing the limitations of existing benchmarks in autonomous driving VLA models [22]. - The dataset is designed to enhance the performance of VLA models in unstructured extreme scenarios, demonstrating significant improvements in established benchmarks [22]. - Experiments show that training with the Impromptu VLA dataset leads to notable performance enhancements in closed-loop NeuroNCAP scores and collision rates [22]. Group 5: DriveMoE - DriveMoE is a novel end-to-end autonomous driving framework that incorporates a mixture-of-experts (MoE) architecture to effectively handle multi-view sensor data and complex driving scenarios [28]. - The framework features scene-specific visual MoE and skill-specific action MoE, addressing the challenges of multi-view redundancy and skill specialization [28]. - DriveMoE achieves state-of-the-art performance in closed-loop evaluations on the Bench2Drive benchmark, demonstrating the effectiveness of combining visual and action MoE in autonomous driving tasks [28].

Core Insights - The article discusses the rapid advancements in end-to-end autonomous driving, particularly focusing on Vision-Language-Action (VLA) models and their applications in the industry [2][3]. Group 1: VLA Model Developments - The introduction of AutoVLA, a new VLA model that integrates reasoning and action generation for end-to-end autonomous driving, shows promising results in semantic reasoning and trajectory planning [3][4]. - ReCogDrive, another VLA model, addresses performance issues in rare and long-tail scenarios by utilizing a three-stage training framework that combines visual language models with diffusion planners [7][9]. - Impromptu VLA introduces a dataset aimed at improving VLA models' performance in unstructured extreme conditions, demonstrating significant performance improvements in established benchmarks [14][24]. Group 2: Experimental Results - AutoVLA achieved competitive performance metrics in various scenarios, with the best-of-N method reaching a PDMS score of 92.12, indicating its effectiveness in planning and execution [5]. - ReCogDrive set a new state-of-the-art PDMS score of 89.6 on the NAVSIM benchmark, showcasing its robustness and safety in driving trajectories [9][10]. - The OpenDriveVLA model demonstrated superior results in open-loop trajectory planning and driving-related question-answering tasks, outperforming previous methods on the nuScenes dataset [28][32]. Group 3: Industry Trends - The article highlights a trend among major automotive manufacturers, such as Li Auto, Xiaomi, and XPeng, to invest heavily in VLA model research and development, indicating a competitive landscape in autonomous driving technology [2][3]. - The integration of large language models (LLMs) with VLA frameworks is becoming a focal point for enhancing decision-making capabilities in autonomous vehicles, as seen in models like ORION and VLM-RL [33][39].