Core Viewpoint - Quark Health's large model has become the first in China to pass the written assessment for chief physician in 12 core medical disciplines, indicating significant advancements in AI healthcare capabilities [1] Group 1: Market Growth and Competition - The global AI healthcare market is projected to grow from $11 billion in 2021 to $194 billion by 2028, with a compound annual growth rate (CAGR) exceeding 41% [1] - Major companies like ByteDance, Baidu, and Alibaba are investing heavily in health large models, highlighting the competitive landscape [1] Group 2: Challenges in Accuracy - The accuracy of health large models remains a core pain point, with challenges including the precision of patient-selected prompts and the development of multimodal capabilities [2] - Accurate understanding of patient expressions and needs is crucial for AI to assist both patients and doctors effectively [2] Group 3: Development of "Slow Thinking" Capability - Quark Health's large model has achieved a breakthrough by developing "slow thinking capability," which integrates chain reasoning and multi-stage clinical deduction to address complex medical issues [2] - High-quality reasoning training data is essential for building this capability, with medical data categorized into "verifiable" and "non-verifiable" types [2] Group 4: Investment in Clinical Data - The development of health large models increasingly relies on clinical data, diagnostics, and data annotation from human doctors [3] - Quark Health has a professional annotation team of over a thousand, including more than 400 senior medical experts [3] Group 5: Commercialization Challenges - Currently, Quark Health is not focusing on commercialization, although future directions may include health record management, diagnostic service conversion, and smart device services [4] - The commercialization of health large models remains a complex issue that is still in the early discussion stages [4]

Core Insights - The current pain point for health large models is insufficient accuracy, as stated by Quark Health's product head [1] - Quark Health's large model has become the first in China to pass the written assessment for chief physicians, following its earlier success with deputy chief physician exams [1] - The global AI in healthcare market is projected to grow from $11 billion in 2021 to $194 billion by 2028, with a compound annual growth rate (CAGR) exceeding 41% [1] Group 1: Challenges and Developments - Health large models face challenges related to the accuracy of consumer-selected prompts and the development of multimodal capabilities, which affect the output of model responses [2] - A significant breakthrough for Quark Health's large model is the development of "slow thinking ability," which integrates chain reasoning and multi-stage clinical reasoning to address complex medical issues [2] Group 2: Training and Commercialization - To build slow thinking ability, high-quality reasoning training data is essential, with Quark categorizing medical data into "verifiable" and "non-verifiable" types for different tasks [5] - Quark Health's large model has a professional annotation team of over 1,000, including more than 400 senior medical experts, highlighting the importance of clinical data and human input for model development [5] - Currently, Quark Health is not considering commercialization, but potential future directions may include health record management and diagnostic service transformations [5]

而为了构建慢思考能力，夸克在研发过程中采用了"双数据产线+双奖励机制"的工程体系。一方面，将 医学数据划分为"可验证"和"不可验证"两类，分别对应诊断类任务和健康建议类任务；另一方面，在训 练方法上引入"过程奖励模型"和"结果奖励模型"，分别评估模型推理链的合理性与最终结论的准确性， 提升模型的临床可解释性和推理一致性。 该体系还设计了多阶段强化学习流程，包括冷启动数据的严格人工校验、多轮样本筛选与难度递进训练 策略，以及用于防止"高分投机"的作弊识别机制。通过真实医生标注、"问—思—答"整组数据驱动强化 学习。 数据显示，夸克健康大模型已拥有千人规模的专业医师标注团队，其中超过400名均为副主任医师及以 上的高资历医疗专家。并且，夸克AI搜索还吸引了一大批医学生和医生群体用户。夸克健康运营负责 人赵存忠透露，目前平台在全国医学生中月活用户已突破200万，覆盖率过半，他们使用夸克用于基础 知识搜索、考试备考和临床辅助诊疗。 7月23日，观察者网获悉，夸克健康大模型成功通过中国12门核心学科的主任医师笔试评测，成为国内 首个完成这一挑战的大模型。目前，"主任级AI医生"能力已全面集成至夸克的AI搜索中，用户在查 ...