Core Viewpoint - The article presents the Embodied Reasoner framework, which extends deep reasoning capabilities to embodied interactive tasks, addressing unique challenges such as multimodal interaction and diverse reasoning patterns [3][7][19]. Group 1: Research Background - Recent advancements in deep reasoning models, like OpenAI's o1, have shown exceptional capabilities in mathematical and programming tasks through large-scale reinforcement learning [7]. - However, the effectiveness of these models in embodied domains requiring continuous interaction with the environment has not been fully explored [7]. - The research aims to expand deep reasoning capabilities to embodied interactive tasks, tackling challenges such as multimodal interaction and diverse reasoning patterns [7]. Group 2: Embodied Interaction Task Design - A high-level planning and reasoning embodied task was designed, focusing on searching for hidden objects in unknown rooms rather than low-level motion control [8]. - The task environment is built on the AI2-THOR simulator, featuring 120 unique indoor scenes and 2100 objects [8]. - Four common tasks were designed: Search, Manipulate, Transport, and Composite Tasks [8]. Group 3: Data Engine and Training Strategy - A data engine was developed to synthesize diverse reasoning processes, presenting embodied reasoning trajectories in an observe-think-act format [3]. - A three-stage iterative training process was introduced, including imitation learning, rejection sampling adjustment, and reflection adjustment, enhancing the model's interaction, exploration, and reflection capabilities [3][19]. - The training corpus synthesized 9390 unique task instructions and their corresponding observe-think-act trajectories, covering 107 different indoor scenes and 2100 interactive objects [12][16]. Group 4: Experimental Results - The model demonstrated significant advantages over existing advanced models, particularly in complex long-duration tasks, showing more consistent reasoning capabilities and efficient search behavior [3][18]. - In real-world experiments, the Embodied Reasoner achieved a success rate of 56.7% across 30 tasks, outperforming OpenAI's o1 and o3-mini [17]. - The model's success rate improved by 9%, 24%, and 13% compared to GPT-o1, GPT-o3-mini, and Claude-3.7-Sonnet-thinking, respectively [18]. Group 5: Conclusion and Future Work - The research successfully extends the deep reasoning paradigm to embodied interactive tasks, demonstrating enhanced interaction and reasoning capabilities, especially in complex long-duration tasks [19]. - Future work may explore the application of the model to a wider variety of embodied tasks and improve its generalization and adaptability in real-world environments [19].

机器之心原创 作者：张倩 对于百度而言，既要保持长期主义的战略定力，也要在技术路径上灵活应变，这种「变与不变」的平衡或许正是其在这轮科技革命中的 制胜之道。 2025 年，模型能力的重要性依然无需多言。 从预训练的角度来看，虽然连 OpenAI 前首席科学家 Ilya Sutskever 都说，预训练数据即将用尽，但海量的图像、视频等多模态数据资源依然有待挖掘。 从后训练的角度来看，强化学习新范式正在让 Scaling Law 焕发新生， 新一代的推理模型在数学、代码、长程规划等问题上不断取得新进展。 对于 AI 公司来说，保持对基础模型研发的投入依然非常必要。现阶段来看，这仍然是攀登智能高峰的本质所在。 而在这个领域，百度一直是一个不可忽视的力量。从 2019 年发布文心大模型 1.0 至今， 文心大模型从知识和数据融合学习，到知识增强、知识点增强，从检索增 强、对话增强、逻辑推理增强，到慢思考、多模态的技术演进并非偶然，而是早期技术探索形成的「积淀」持续推动的结果。 正是这份「积淀」引领百度打造出 超越 GPT-4o 的多模态大模型文心 4.5 Turbo，以及领先 DeepSeek R1、V3 的深度思 ...

Core Viewpoint - The development direction of the industry is to create Agents with advanced reflection, planning, and autonomous decision-making capabilities, moving beyond the current basic level of many Agents [1] Group 1: Agent Development - The foundation for building Agents is deep thinking models, which must possess the ability to think, plan, and reflect [1] - Agents should support multimodal capabilities, similar to human visual and auditory functions, to better handle complex tasks [1] Group 2: Product Launch - The Doubao 1.5 deep thinking model was officially launched, showcasing strong performance in general tasks such as mathematics, programming, scientific reasoning, and creative writing [1] - A visual version of the deep thinking model was introduced, which has visual reasoning capabilities, allowing it to associate and think about what it sees like a human [1]

备受瞩目的 2025 全球机器学习技术大会（ML Summit 2025）将于 4 月 18-19 日在上海虹桥西郊庄园丽笙大酒店召开。本次盛会由 CSDN & Boolan 联合主办，汇聚了超 50 位来自学术界和工业界顶尖专家，共同探讨智能体、联邦学习、多模态大模型等热门 AI 技术实践。 作为全球机器学习技术大会的老朋友，新浪微博首席科学家及 AI 研发部负责人张俊林将带来《从 DeepSeek R1 的复现看深度思考模型的未来》的精 彩分享。 张俊林作为「大模型技术拆解得最通透的实战派」，在 2024 年的机器学习技术大会上，他对 Gemini 多模态架构、OpenAI o1 技术的硬核拆解，让 开发者直呼"终于有人讲透技术本质"。 系统梳理技术脉络： 回顾 DeepSeek R1 开源后的各类复现研究，涵盖 SFT 阶段的轻量适配（如 S1）与 RL 阶段的创新实践。 深度解析训练范式： 重点剖析其核心的两阶段训练模式——如何通过冷启动微调结合多领域数据优化进行 SFT，以及如何运用 GRPO 强化学习 与全场景对齐实现模型"深度思考"能力的跃迁。 探讨关键技术问题： 尝试解答一系列备受关注的核心问 ...