CoT

Search documents
只用2700万参数,这个推理模型超越了DeepSeek和Claude
机器之心· 2025-06-30 10:23
机器之心报道 编辑:泽南、陈陈 像人一样推理。 大模型的架构,到了需要变革的时候? 在对复杂任务的推理工作上,当前的大语言模型(LLM)主要采用思维链(CoT)技术,但这些技术存在任务分解复杂、数据需求大以及高延迟等问题。 近日,受到人脑分层和多时间尺度处理机制启发,来自 Sapient Intelligence 的研究者提出了分层推理模型(HRM),这是一种全新循环架构,能够在保持训练稳定 性和效率的同时,实现高计算深度。 具体来说,HRM 通过两个相互依赖的循环模块,在单次前向传递中执行顺序推理任务,而无需对中间过程进行明确的监督:其中一个高级模块负责缓慢、抽象的 规划,另一个低级模块负责处理快速、细致的计算。HRM 仅包含 2700 万个参数,仅使用 1000 个训练样本,便在复杂的推理任务上取得了卓越的性能。 该模型无需预训练或 CoT 数据即可运行,但在包括复杂数独谜题和大型迷宫中最优路径查找在内的挑战性任务上却取得了近乎完美的性能。此外,在抽象与推理 语料库 (ARC) 上,HRM 的表现优于上下文窗口明显更长的大型模型。ARC 是衡量通用人工智能能力的关键基准。 由此观之,HRM 具有推动通用计 ...
X @The Wall Street Journal
The Wall Street Journal· 2025-06-29 23:00
Canadians’ boycott of a beloved seaside town in Maine this summer feels more like a heart-wrenching breakup https://t.co/yQc3FTSLVz ...
X @The Wall Street Journal
The Wall Street Journal· 2025-06-29 19:12
Canadians’ boycott of a beloved seaside town in Maine this summer feels more like a heart-wrenching breakup https://t.co/HF2dml7Wyq ...
豆包1.6 “不偏科” ,高考成绩直逼“清北”
2 1 Shi Ji Jing Ji Bao Dao· 2025-06-28 14:29
Core Insights - The Seed1.6-Thinking model from Doubao achieved impressive scores in the 2025 college entrance examination, with a total score of 683 in liberal arts and 648 in sciences, indicating its strong performance across various subjects [1][2] - The model's results suggest it is competitive enough to potentially gain admission to top universities like Tsinghua and Peking University, with predictions indicating a possible score exceeding 690 in key subjects [2][3] Performance Summary - Doubao's Seed1.6-Thinking model excelled in multiple subjects, achieving the highest scores in Chinese, English, Physics, History, Geography, and Politics, with a mathematics score exceeding 140 [2] - In an international test, the model also ranked among the top performers in the JEE Advanced exam in India, showcasing its capabilities in mathematics, physics, and chemistry [3] Model Capabilities - The Seed team clarified that the model does not exhibit a bias towards specific subjects, as it demonstrated improved performance in chemistry and biology after using higher-quality test images [4] - The introduction of "dynamic thinking ability" (AutoCoT) allows the model to adapt its reasoning process, enhancing its performance while reducing unnecessary complexity in reasoning [4][6] Industry Implications - The potential of AI in the education sector, particularly in college entrance examinations, has garnered attention, with AI tools being developed to assist in decision-making for college applications [5] - The Seed1.6 model represents a significant advancement in AI capabilities, integrating multimodal understanding and deep reasoning, and is now available for API access through Volcano Engine [6]
Achieve Life Sciences Announces Proposed Underwritten Public Offering
Globenewswire· 2025-06-26 20:03
Company Overview - Achieve Life Sciences, Inc. is a late-stage specialty pharmaceutical company focused on the global development and commercialization of cytisinicline as a treatment for nicotine dependence and smoking cessation [5] - The company submitted its New Drug Application to the FDA for cytisinicline in June 2025, based on two completed Phase 3 studies and a fully enrolled open-label safety study [5] Offering Details - Achieve Life Sciences announced a proposed underwritten public offering to sell shares of its common stock and accompanying common warrants, with an option for underwriters to purchase an additional 15% of the shares [1][2] - The proceeds from the offering are intended to fund the advancement of cytisinicline through potential FDA marketing approval and for working capital and general corporate purposes [2] Market Context - Approximately 29 million adults in the U.S. smoke combustible cigarettes, with tobacco use being the leading cause of preventable death, responsible for over eight million deaths globally and nearly half a million in the U.S. annually [6] - There are around 17 million adults in the U.S. who use e-cigarettes, with no FDA-approved treatments specifically for nicotine e-cigarette cessation, highlighting a critical need for effective solutions [7] Product Information - Cytisinicline is a plant-based alkaloid with a high binding affinity to nicotinic acetylcholine receptors, believed to aid in treating nicotine addiction by reducing craving symptoms and the satisfaction associated with nicotine products [8][9] - Cytisinicline has been granted Breakthrough Therapy designation by the FDA to address the urgent need for treatments for nicotine dependence [7]
Achieve Life Sciences Announces Submission of NDA to FDA for Cytisinicline as a Treatment of Nicotine Dependence for Smoking Cessation
Globenewswire· 2025-06-26 20:01
Core Insights - Achieve Life Sciences has submitted a New Drug Application (NDA) to the FDA for cytisinicline, marking the first new pharmacotherapy option for nicotine dependence in two decades [1][4] - Cytisinicline has shown efficacy and safety in two large Phase 3 trials, ORCA-2 and ORCA-3, demonstrating significantly higher abstinence rates compared to placebo [2][3] - The public health burden of smoking affects nearly 29 million adults in the U.S., with smoking-related illnesses causing nearly half a million deaths annually [2][5] Company Overview - Achieve Life Sciences is a late-stage specialty pharmaceutical company focused on developing cytisinicline for nicotine dependence treatment [4] - The company has also completed a Phase 2 study for vaping cessation and has received Breakthrough Therapy designation from the FDA for cytisinicline [6][4] Product Details - Cytisinicline is a plant-based alkaloid that interacts with nicotinic acetylcholine receptors, potentially reducing nicotine cravings and satisfaction from nicotine products [7] - The NDA submission is supported by data from over 2,000 clinical trial participants, indicating a well-tolerated safety profile [1][2]
国产大模型高考出分了:裸分683,选清华还是北大?
量子位· 2025-06-26 06:25
Core Insights - The article discusses the performance of various AI models in a simulated high school examination, comparing their scores and capabilities in different subjects [2][12]. Group 1: Overall Performance - Gemini achieved the highest score in science with 655 points, while Doubao scored 683 points in humanities, also ranking first [2]. - Doubao excelled in six subjects, maintaining top scores except in mathematics, chemistry, and biology [3][4]. Group 2: Subject-Specific Analysis - In the subject breakdown, Doubao scored 128 in Chinese, 141 in mathematics, and 144 in English, while Gemini scored 126 in Chinese and 140 in mathematics [3]. - The models showed significant improvement in mathematics compared to previous years, with most scoring around 140 points [13]. - Doubao and Gemini demonstrated better performance in visual comprehension tasks compared to other models, particularly in chemistry [22][42]. Group 3: Evaluation Methodology - The evaluation used a combination of national and provincial exam papers, with a total score of 750 points [9]. - Scoring was conducted through a mix of automated assessments and human evaluations, ensuring a fair testing environment [10][11]. Group 4: Model Development and Improvement - Doubao's advancements are attributed to three key strategies: multi-modal integration, enhanced reasoning capabilities, and dynamic thinking abilities [30][33][40]. - The model's training involved a three-phase process focusing on text, multi-modal data, and long-context support, significantly improving its performance in reading comprehension and reasoning tasks [35][36]. Group 5: Future Directions - The article suggests that combining text and image inputs can significantly enhance model performance, indicating a promising area for future exploration [42][43].
22nd Century Announces Second Partner VLN Product Deal as Part of Major Pinnacle Brand Expansion Agreement with Top-5 C-Store Chain
Globenewswire· 2025-06-24 13:00
Core Insights - 22nd Century Group, Inc. is launching new Pinnacle VLN and moist snuff products in over 1,700 stores across 27 states, marking a significant expansion in its product offerings [1][2] - The new products include Pinnacle VLN Gold and Menthol VLN cigarettes, which are expected to begin sales in late summer and early fall of 2025 [2][4] - The company aims to leverage its established Pinnacle brand, which has a strong sales track record, to drive success in the new product categories [2][5] Product Launch Details - The launch includes four new Pinnacle SKUs, with two specifically in the low nicotine category [1] - The moist snuff products will feature straight and wintergreen flavors, expected to be available in the second half of 2025 [4] - The manufacturing of these products will utilize proprietary VLN tobacco strains and will be distributed through existing national-scale distribution agreements [5][6] Company Background - 22nd Century Group is recognized as a pioneering nicotine harm reduction company, focusing on enabling smokers to control their nicotine consumption [7] - The flagship VLN cigarette contains 95% less nicotine than traditional cigarettes, providing an alternative for smokers looking to reduce their nicotine intake [8][10] - The company operates a facility in Mocksville, North Carolina, with the capacity to produce over 45 million cartons of combustible tobacco products annually [9]
瑞达期货棉花(纱)产业日报-20250623
Rui Da Qi Huo· 2025-06-23 11:20
外谨慎,仅根据实际生产需求补充库存。消费淡季,去库存速度缓慢,旧作基本面变化较小,短期价格震 数据来源第三方(wind、同花顺、棉花信息网、棉花协会网),观点仅供参考。市场有风险,投资需谨慎! 荡。近期全疆大部处于现蕾开花期的棉花存在不同程度高温热害风险,持续关注主产区新季棉花生长情况 研究员: 张昕 期货从业资格号F03109641 期货投资咨询从业证书号Z0018457 对盘面影响。 免责声明 本报告中的信息均来源于公开可获得资料,瑞达期货股份有限公司力求准确可靠,但对这些信息的准确性及完整性不做任 何保证,据此投资,责任自负。本报告不构成个人投资建议,客户应考虑本报告中的任何意见或建议是否符合其特定状况。本 报告版权仅为我公司所有,未经书面许可,任何机构和个人不得以任何形式翻版、复制和发布。如引用、刊发,需注明出处为 瑞达期货股份有限公司研究院,且不得对本报告进行有悖原意的引用、删节和修改。 棉花(纱)产业日报 2025-06-23 | 项目类别 | 数据指标 | 最新 | 环比 数据指标 | 最新 | 环比 | | --- | --- | --- | --- | --- | --- | | 期货市场 ...
细粒度视觉推理链引入数学领域,准确率暴涨32%,港中文MMLab打破多模态数学推理瓶颈
量子位· 2025-06-16 10:30
MINT-CoT团队 投稿 量子位 | 公众号 QbitAI 思维链(Chain of Thought, CoT)推理方法已被证明能够显著提升大语言模型(LLMs)在复杂任务中的表现。而在多模态大语言模型 (MLLMs)中,CoT 同样展现出了巨大潜力。 3. 过度依赖外部功能 像 MVoT 或 Visual SKETCHPAD 等方法,需要借助外部工具或能力来生成或修改图像,训练和推理过程成本高、不通用。 然而,当视觉信息与数学推理结合时,传统的 CoT 方法就显得力不从心了——视觉输入中的数学细节往往被忽略,导致推理结果不准确。 最近,香港中文大学 MMLab 团队正式发布了全新的视觉推理方案——MINT-CoT,专为解决"多模态数学推理"中的难题而设计。 为什么数学视觉推理这么难? 尽管已有一些研究尝试把视觉信息引入 CoT 推理,例如 Visual-CoT、Visual SKETCHPAD、VPT、ICoT 等方法,但在数学场景下依然存 在 三大瓶颈: 1. 粗粒度图像区域选择 大部分方法依赖边界框(Bounding Box)来截取图像区域。但数学图像里的元素(比如坐标轴、几何图形、标注文字等)高度关 ...