[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f_MEqH-K-mkmCPCoF1ZdZum2nP57Md76nG0VaOAu260o":3},{"code":4,"msg":5,"data":6},200,"操作成功",{"id":7,"title":8,"content":9,"digest":10,"source":10,"coverPath":11,"thumbsCoverPath":12,"isTop":13,"isShow":14,"baseClick":13,"clickCount":15,"createTime":16,"typeId":17,"isNewest":18,"newsInfoTypeRespVo":19,"voiceUrl":22,"voiceSize":23,"taskId":24,"releaseTime":25,"titleEn":26,"contentEn":27,"voiceUrlEn":28,"taskIdEn":29,"voiceSizeEn":30},1186,"AMD发布SAND-Math：用AI造题神器让数学学习更有趣，解决理工科教育资源稀缺难题","\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F44697899bc2b4ec5bdb186fbc05257df\u002FAA1JUH1Z.png\" width=\"624\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\">这项由AMD公司高级研究团队完成的突破性研究于2025年7月发表，论文标题为\"SAND-Math: Using LLMs to Generate Novel, Difficult and Useful Mathematics Questions and Answers\"。该研究由AMD的五位资深研究员Chaitanya Manem、Pratik Prabhanjan Brahma、Prakamya Mishra、Zicheng Liu和Emad Barsoum共同完成。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">（完整论文已在arXiv平台发布，论文编号为2507.20527v1，有兴趣深入了解的读者可以通过https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002Famd\u002FSAND-MATH访问相关数据集。）\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在当今这个AI技术飞速发展的时代，我们经常听说AI能写文章、画图片，甚至能编程序，但你可能没想到，AI现在还能当数学老师，专门出那些让学生\"又爱又恨\"的数学题。AMD的研究团队刚刚发布了一项令人兴奋的成果，他们开发出了一套名为SAND-Math的系统，这套系统就像一个永不疲倦的数学题库生成器，能够源源不断地创造出新颖、困难且实用的数学问题。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">要理解这项研究的重要性，我们不妨先回到现实生活中。每一位数学老师都面临着这样的困扰：如何找到足够多、足够好的数学题来训练学生？传统的做法是从各种数学竞赛、奥林匹克数学题库中搜集题目，但这些题目数量有限，而且往往重复性较高。更重要的是，随着AI技术的发展，我们需要更聪明的AI来帮助学生学习数学，但训练这些AI需要大量高质量的数学问题和解答，而现有的数学题库远远无法满足这种需求。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">AMD的研究团队敏锐地察觉到了这个问题。他们发现，虽然市面上已有一些AI生成数学题的方法，但这些方法就像在旧题目上\"换汤不换药\"，生成的新题目往往只是把原题中的数字改一改，或者稍微调整一下题目背景，本质上仍然是同一类型的问题。这就好比一个厨师只会做蛋炒饭，无论怎么变化，都逃不出\"炒饭\"的范畴。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">SAND-Math系统的出现就像给这个厨师提供了一整套全新的烹饪工具和食材。这套系统不是简单地改造现有题目，而是从零开始创造全新的数学问题。更神奇的是，它还配备了一个\"难度调节器\"，能够系统性地提高问题的复杂程度，就像游戏中的关卡设计师，能够精确控制每一关的挑战难度。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">研究团队采用了一种非常巧妙的方法。他们认为，最先进的AI模型（比如那些能够解决复杂数学问题的大语言模型）实际上已经具备了一种\"数学直觉\"，能够理解什么样的问题是困难的，什么样的问题是有意义的。基于这个假设，他们设计了一套完整的数学题生成流水线，这个流水线就像一个高度自动化的工厂，能够批量生产高质量的数学问题。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">一、SAND-Math系统的工作原理：像搭积木一样造数学题\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">SAND-Math系统的工作流程可以想象成一个精密的质量控制工厂。这个工厂的目标是生产出既新颖又困难、同时还具有教育价值的数学题目。整个生产过程分为多个环节，每个环节都有严格的质量检查，确保最终产品符合要求。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">工厂的第一个车间是\"创意生成部\"。在这里，系统会向AI模型提出一个看似简单的要求：\"请生成一道奥林匹克水平的数学题\"。这个请求看似简单，但实际上激发了AI模型内在的数学知识储备。就像一位经验丰富的数学老师，即使只给他一个简单的提示，他也能凭借多年的教学经验创造出有挑战性的题目。系统通过这种方式，在不依赖任何现有题库的情况下，从头开始生成了23,437道原创数学题。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">接下来，这些新生成的题目会进入\"答案验证车间\"。在这里，系统会让AI模型从不同角度、用不同方法来解决同一道题目。这个过程就像让多位数学老师独立解答同一道题，然后比较他们的答案是否一致。只有当所有解答方法都得出相同答案的题目，才能通过这一关的检验。这种做法确保了题目的准确性和可解性，避免了那些无解或答案错误的问题混入最终的题库。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">通过答案验证的题目接着会来到\"原创性检查站\"。系统会使用先进的文本相似度检测技术，将每道新题目与互联网上的海量数学内容进行比较，确保没有抄袭或重复现有的题目。这个过程就像学术论文的查重检测，但应用在数学题目上。令人惊喜的是，经过检查后发现，通过SAND-Math系统生成的题目几乎都是全新的，重复率极低，这证明了系统确实具备了创造原创数学问题的能力。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">最有趣的是\"难度筛选环节\"。研究团队使用了一个聪明的策略：他们让一个相当强大的AI模型（Qwen2.5-32B-Instruct）来尝试解答这些题目。凡是这个\"考官\"无法正确解答的题目，就被认为具有足够的挑战性，可以保留下来用于训练更高级的AI系统。这就像用一个优秀学生来帮忙筛选题目，那些连优秀学生都觉得困难的题目，显然更适合用来挑战和提升AI的数学能力。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">二、难度攀登技术：让简单题目变身数学难题\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">SAND-Math系统最独特的创新是它的\"难度攀登\"（Difficulty Hiking）技术。这项技术可以比作一个数学题的\"升级器\"，能够将相对简单的题目系统性地改造成更具挑战性的高难度问题。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">整个难度攀登过程就像一个经验丰富的数学老师在备课时的思考过程。当老师拿到一道基础题目后，他会思考：如何在保持题目核心数学思想的同时，增加更多的数学概念和推理步骤，让题目变得更加复杂和有挑战性？\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">系统的做法是为每道原始题目配备四个关键要素：首先是题目本身及其解答过程，这提供了基础的数学框架；然后是当前的难度评级，这让系统知道题目的起始难度水平；接着系统会从数学知识库中选择一个相关的高级定理，这个定理必须与原题目属于同一数学分支，确保数学逻辑的连贯性；最后，系统会随机选择一个来自其他数学分支的概念或工具，用于增加题目的跨领域复杂性。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">有了这四个要素后，系统会指导AI模型将它们巧妙地融合在一起，创造出一道全新的高难度题目。这个过程的巧妙之处在于，新题目必须自然地整合所有这些元素，而不是简单地把它们拼凑在一起。就像一位大厨不会简单地把各种食材混合，而是要让它们在口味和营养上形成完美的搭配。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">研究团队测试了这种难度攀登技术的效果，结果令人印象深刻。经过一轮难度攀登处理，题目的平均难度评分从5.02分提升到了5.98分（满分10分）。更重要的是，评分在6.0分以上的困难题目比例从47.2%大幅增加到76.8%。这种变化不仅体现在数字上，更体现在实际应用效果中：用经过难度攀登处理的题目训练的AI模型，在数学推理能力测试中的表现明显优于使用原始题目训练的模型。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">为了更好地理解这个过程，我们可以看一个具体例子。假设原始题目是一个关于三角函数的相对简单问题，系统可能会引入复分析中的高级定理，同时融入组合数学中的概念，最终生成一道需要同时运用三角函数、复分析和组合思维的综合性难题。这样的题目不仅保持了原题的数学精神，还大大增加了解题的复杂性和思维深度。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">三、实验验证：用数据说话的教学效果\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">为了验证SAND-Math系统的实际效果，研究团队进行了一系列对比实验，就像进行教学效果的对照研究一样。他们选择了Qwen2.5-32B这个强大的AI模型作为\"学生\"，然后用不同的数学题集来训练它，观察哪种训练方式能让这个AI\"学生\"在数学考试中表现得更好。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">实验的设计非常巧妙。研究团队首先建立了一个基准测试，使用LIMO数据集（一个已知的高质量数学题库）来训练AI模型。这就像给学生提供一本经典的数学练习册。然后，他们分别用SAND-Math生成的题目、其他现有的数学题库，以及两者的组合来训练同样的AI模型，观察不同训练方式对模型数学能力的影响。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">评测标准选择了几个具有代表性的数学竞赛：2024年和2025年的美国数学邀请赛（AIME）、美国数学竞赛（AMC），以及MATH-500测试集。这些测试就像不同类型的数学考试，能够从多个角度评估AI模型的数学推理能力。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">实验结果让人眼前一亮。当将SAND-Math的题目作为补充训练材料时，AI模型在AIME25测试中的表现从基准的71.50%提升到了73.32%。更令人印象深刻的是，这个提升幅度比使用其他任何现有数学题库都要大。具体来说，SAND-Math的提升效果比次好的合成数据集高出了17.85个百分点，这在AI训练领域是一个相当显著的改进。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">研究团队还专门测试了难度攀登技术的独立效果。他们发现，使用经过难度攀登处理的题目训练的模型，比使用原始题目训练的模型表现更好。在AIME25测试中，难度攀登技术将模型的得分从46.38%提升到了49.23%。这个结果清楚地表明，不是题目越多越好，而是题目越有挑战性，训练效果越好。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这些实验结果的意义不仅仅在于数字上的提升。\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\">它们证明了一个重要的教育理念：高质量的练习材料比大量的重复练习更能提升学习效果。\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">就像体能训练中的\"渐进式超负荷\"原理一样，只有不断增加挑战难度，才能真正提升能力水平。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">四、技术实现细节：构建一个数学题工厂的技术秘密\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">SAND-Math系统的技术实现就像搭建一个高度自动化的工厂，每个环节都需要精确的技术配置和质量控制机制。研究团队在技术选择和系统设计上投入了大量心血，确保整个流程既高效又可靠。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">系统的核心引擎选择了DeepSeek-R1模型作为主要的\"数学题创作者\"。这个选择并非随意，而是基于该模型在数学推理方面的出色表现。在题目生成阶段，系统将模型的创造性参数（temperature）设置为0.8，这个数值就像调节创意的旋钮，既保证了生成内容的多样性，又避免了过于随机的输出。在解答生成阶段，参数被调整为0.6，稍微降低随机性以确保解答的准确性。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">质量控制环节采用了Llama-3.3-70B-Instruct模型作为\"质检员\"。这个模型负责验证答案的一致性、识别重复内容、评估题目难度。为了提高评估的可靠性，每道题目的难度评分都要进行3次独立评估，然后取平均值。这种做法就像让多位专家独立打分，最后综合评判，大大提高了评估结果的客观性。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">去重和查重环节使用了先进的语义哈希技术。系统采用semhash框架配合minisilab\u002Fpotion-base-8M模型，在0.99的相似度阈值下检测内容重复。这个过程就像图书馆的查重系统，能够识别出那些表述不同但本质相同的题目。有趣的是，系统还配备了网络搜索功能，会将每道生成的题目作为搜索关键词，在互联网上寻找相似内容，确保生成的题目确实是原创的。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">整个系统的硬件配置也颇为豪华：8块AMD InstinctTM MI300X GPU构成了计算集群，所有模型都在单一节点上运行，这样的配置确保了处理速度和稳定性。为了优化性能，团队还采用了DeepSpeed框架的ZeRO-3技术进行内存管理，这些技术细节虽然对普通用户来说比较抽象，但它们共同保证了系统能够高效地处理大量数学题目的生成和筛选工作。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">值得一提的是，研究团队还开发了一套专门的数学知识分类体系。他们将数学知识细分为数论、代数、几何、组合数学、概率论等多个分支，每个分支下又包含了详细的定理和概念清单。这个知识体系就像一个巨大的数学百科全书，为难度攀登技术提供了丰富的素材库。当系统需要为某道题目增加难度时，它会从这个知识库中智能选择相关的定理和概念，确保生成的新题目在数学上是合理和有意义的。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">五、实际应用与未来展望：数学教育的革新之路\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">SAND-Math系统的出现为数学教育带来了全新的可能性。在传统的教学模式中，老师们往往需要花费大量时间搜集和整理数学题目，而且受限于现有题库的数量和质量。\u003C\u002Fstrong>\u003Cstrong style=\"font-size: 18px;\">现在，有了这样一个能够无限生成高质量数学题的系统，教育工作者可以根据学生的具体需求，定制不同难度和类型的练习材料。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这种技术最直接的应用场景是个性化学习。每个学生的数学基础和学习进度都不相同，传统的\"一刀切\"教学方式往往无法满足所有学生的需求。而SAND-Math系统可以根据学生的当前水平，生成恰到好处的练习题目。对于基础较弱的学生，系统可以生成更多基础性题目；对于学有余力的学生，系统可以提供更具挑战性的问题。这就像为每个学生量身定制了一套专属的数学练习册。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">另一个重要应用是AI教育助手的训练。随着人工智能在教育领域的应用越来越广泛，我们需要更聪明、更能干的AI来辅助数学教学。SAND-Math生成的高质量题库为训练这些AI教育助手提供了宝贵的资源。通过在这些精心设计的题目上进行训练，AI助手能够更好地理解数学概念，提供更准确的解题指导。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">从更宏观的角度来看，这项技术还可能改变数学竞赛和考试的命题方式。传统的命题工作主要依赖专家的经验和灵感，不仅耗时耗力，而且容易出现风格单一的问题。SAND-Math系统可以作为命题专家的得力助手，快速生成大量候选题目，然后由专家进行筛选和完善。这种人机协作的模式既保证了题目的质量，又大大提高了命题效率。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">当然，任何技术都有其局限性。SAND-Math系统生成的题目质量很大程度上依赖于训练模型的能力。如果\"老师\"模型本身存在知识盲区或偏见，生成的题目也可能存在相应的问题。另外，系统目前主要关注题目的数学难度和新颖性，但对于题目的教育价值和实际应用背景的考虑还有提升空间。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">研究团队也坦承，他们的实验主要基于相对较小的样本进行，虽然结果令人鼓舞，但要想充分发挥系统的潜力，还需要在更大规模的数据集上进行验证。他们计划在未来的研究中扩大实验规模，同时探索将这种技术应用到其他学科领域的可能性。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">从长远来看，SAND-Math代表的不仅仅是一个数学题生成工具，更是教育技术发展的一个重要里程碑。它展示了人工智能如何能够创造性地解决教育资源稀缺的问题，为每个学习者提供更丰富、更个性化的学习体验。随着技术的不断完善和应用场景的扩展，我们有理由相信，这样的智能教育工具将在未来的教育体系中发挥越来越重要的作用。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">说到底，SAND-Math系统的真正价值不在于它能生成多少道数学题，而在于它为教育公平和个性化学习提供了新的可能性。当每个学生都能获得适合自己水平的练习材料，当每位老师都能轻松获得高质量的教学资源，我们的数学教育将变得更加高效和有趣。这项由AMD团队开发的技术，正在悄悄地改变着数学学习的面貌，让这门抽象而美丽的学科变得更加亲近和可及。对于那些希望深入了解技术细节的读者，完整的研究论文和相关数据集都已在网络上公开发布，为进一步的研究和应用提供了坚实的基础。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Q&A\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">Q1：SAND-Math系统到底是什么？它能解决什么问题？\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A：SAND-Math是AMD开发的AI数学题生成系统，就像一个永不疲倦的数学老师，能从零开始创造全新的高质量数学题。它主要解决当前数学教育中优质题目稀缺的问题，特别是训练AI数学助手时缺乏足够多样化、有挑战性题目的困扰。系统不仅能生成题目，还能自动调节难度，为不同水平的学习者提供合适的练习材料。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">Q2：这个系统生成的数学题质量怎么样？会不会出错？\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A：系统有严格的质量控制机制。每道题目都要经过多轮验证：先让AI从不同角度解答同一题目，只有答案完全一致的才能通过；然后检查是否与现有题目重复；最后用高水平AI模型测试难度。实验显示，用SAND-Math题目训练的AI在数学测试中比其他方法高出17.85个百分点，证明了题目的高质量。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">Q3：普通老师和学生现在能用上这个系统吗？\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A：目前系统主要用于研究和AI模型训练，还不是面向普通用户的产品。不过研究团队已经公开了相关数据集和技术细节，这为开发面向教育工作者的应用奠定了基础。未来很可能会出现基于这项技术的教学辅助工具，让老师能够根据学生需求定制练习题目。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(187, 187, 187);\">【新闻来源】科技行者 \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Famd%E5%8F%91%E5%B8%83sand-math-%E7%94%A8ai%E9%80%A0%E9%A2%98%E7%A5%9E%E5%99%A8%E8%AE%A9%E6%95%B0%E5%AD%A6%E5%AD%A6%E4%B9%A0%E6%9B%B4%E6%9C%89%E8%B6%A3-%E8%A7%A3%E5%86%B3%E7%90%86%E5%B7%A5%E7%A7%91%E6%95%99%E8%82%B2%E8%B5%84%E6%BA%90%E7%A8%80%E7%BC%BA%E9%9A%BE%E9%A2%98\u002Far-AA1JUS3R?ocid=BingNewsLanding&cvid=482a3bbb03ef4b21cd85ccb8b6128d11&ei=28\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">http:\u002F\u002Fu5a.cn\u002FRQoEQ\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（本网转发此文章，旨在为读者提供更多的信息资讯，所涉内容不构成投资、消费建议。文章事实如有疑问，请与有关方核实，文章观点非本网观点，仅供读者参考。）\u003C\u002Fspan>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F69c5fbf3995a47329dcf5005f66608e9\u002F教育生态.jpg","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002Fthumbs\u002F69c5fbf3995a47329dcf5005f66608e9\u002F教育生态.jpg",0,1,234,"2025-08-06 17:39",2,false,{"id":17,"name":20,"enName":21},"芯位视野","Xinwei Vision","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A00da133d-e4bd-4013-8426-7d859b5926b3%3A0.wav?Expires=1754479961&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=Cv0m%2FPp%2B2ltqOgn%2BEcjS4DUfejI%3D",31392588,"00da133d-e4bd-4013-8426-7d859b5926b3","2025-08-06 17:27","AMD releases SAND-Math: An AI Question-Generating Tool to Make Math Learning More Interesting and Solve the Problem of Scarce Science and Engineering Education Resources","\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F44697899bc2b4ec5bdb186fbc05257df\u002FAA1JUH1Z.png\" width=\"624\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\">This breakthrough research completed by AMD's senior research team was published in July 2025, with the paper title \"SAND-Math: Using LLMs to Generate Novel, Difficult and Useful Mathematics Questions and Answers\". The study was jointly completed by five senior researchers from AMD: Chaitanya Manem, Pratik Prabhanjan Brahma, Prakamya Mishra, Zicheng Liu, and Emad Barsoum.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">(The full paper has been published on the arXiv platform, with the paper number 2507.20527v1. Readers interested in gaining a deeper understanding can access the related dataset at https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002Famd\u002FSAND-MATH.)\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In today's era of rapid development of AI technology, we often hear that AI can write articles, draw pictures, and even code programs, but you might not have expected that AI can now act as a math teacher, specifically creating math problems that students \"love and hate.\" AMD's research team has just released an exciting result, developing a system called SAND-Math, which acts like an endless math question bank generator, capable of continuously creating novel, difficult, and practical math problems.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">To understand the importance of this research, let's go back to real life. Every math teacher faces the problem of how to find enough and good math problems to train students? Traditional methods involve collecting questions from various math competitions and Olympiad math question banks, but these questions are limited in quantity and often repetitive. More importantly, with the development of AI technology, we need smarter AI to help students learn math, but training such AI requires a large amount of high-quality math problems and answers, and existing math question banks are far from meeting this demand.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">AMD's research team noticed this issue keenly. They found that although there are already some AI-generated math question methods available on the market, these methods are like \"changing the soup but not the medicine,\" generating new questions that are often just changing the numbers in the original questions or slightly adjusting the question background, essentially still the same type of problem. This is like a chef who only knows how to cook fried rice, no matter how they change it, they never get out of the \"fried rice\" category.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The emergence of the SAND-Math system is like providing this chef with a complete set of new cooking tools and ingredients. This system does not simply modify existing questions but creates entirely new math problems from scratch. Even more miraculously, it also comes with a \"difficulty regulator,\" which can systematically increase the complexity of the problems, like a game level designer who can precisely control the challenge difficulty of each level.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The research team used a very clever method. They believe that the most advanced AI models (such as those that can solve complex mathematical problems) actually already have a \"mathematical intuition,\" able to understand what kind of problems are difficult and what kind of problems are meaningful. Based on this assumption, they designed a complete math question generation pipeline, which is like a highly automated factory, capable of mass-producing high-quality math problems.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">I. How the SAND-Math System Works: Creating Math Problems Like Building Blocks\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The workflow of the SAND-Math system can be imagined as a precision quality control factory. This factory aims to produce math problems that are both novel and difficult, and also have educational value. The entire production process is divided into multiple stages, each with strict quality checks to ensure that the final product meets the requirements.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The first workshop of the factory is the \"Creative Generation Department.\" Here, the system will ask the AI model for a seemingly simple request: \"Please generate a math problem at the Olympiad level.\" This request seems simple, but it actually triggers the AI model's internal mathematical knowledge reserve. Just like an experienced math teacher, even if given only a simple hint, he can create challenging questions based on years of teaching experience. The system generates 23,437 original math problems from scratch without relying on any existing question bank through this method.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Next, these newly generated questions will enter the \"Answer Verification Workshop.\" Here, the system will ask the AI model to solve the same question from different angles and using different methods. This process is like having multiple math teachers independently solve the same question and then compare their answers to see if they are consistent. Only questions that yield the same answer through all solving methods can pass this stage of inspection. This approach ensures the accuracy and solvability of the questions, avoiding the inclusion of unsolvable or incorrect questions in the final question bank.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Questions that pass the answer verification will then arrive at the \"Originality Check Station.\" The system uses advanced text similarity detection technology to compare each new question with the vast amount of math content on the internet to ensure there is no plagiarism or repetition of existing questions. This process is like academic paper plagiarism detection, but applied to math questions. Surprisingly, after checking, it was found that the questions generated by the SAND-Math system are almost all new, with extremely low duplication rates, proving that the system indeed has the ability to create original math problems.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The most interesting part is the \"Difficulty Screening Stage.\" The research team used a clever strategy: they asked a fairly powerful AI model (Qwen2.5-32B-Instruct) to try to solve these questions. Any questions that this \"examiner\" cannot correctly solve are considered to have sufficient challenge and are retained for training more advanced AI systems. This is like using an excellent student to help screen questions, and those questions that even excellent students find difficult are obviously more suitable for challenging and improving the math capabilities of AI.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">II. Difficulty Climbing Technology: Transforming Simple Questions into Math Challenges\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The most unique innovation of the SAND-Math system is its \"Difficulty Climbing\" (Difficulty Hiking) technology. This technology can be compared to a \"upgrader\" for math questions, capable of systematically transforming relatively simple questions into more challenging high-level problems.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The entire difficulty climbing process is like the thinking process of an experienced math teacher when preparing lessons. When a teacher gets a basic question, he thinks: How can I add more mathematical concepts and reasoning steps while maintaining the core mathematical idea of the question to make it more complex and challenging?\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">The system's approach is to equip each original question with four key elements: first, the question itself and its solution process, which provides the basic mathematical framework; then, the current difficulty rating, which lets the system know the starting difficulty level of the question; next, the system selects a relevant advanced theorem from the mathematical knowledge base, which must belong to the same mathematical branch as the original question to ensure the continuity of mathematical logic; finally, the system randomly selects a concept or tool from another mathematical branch to increase the cross-disciplinary complexity of the question.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">With these four elements, the system guides the AI model to skillfully integrate them to create a brand-new high-difficulty question. The ingenuity of this process lies in the fact that the new question must naturally integrate all these elements, rather than simply putting them together. It is like a master chef who doesn't simply mix various ingredients, but rather makes them form a perfect match in taste and nutrition.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The research team tested the effectiveness of this difficulty climbing technology, and the results were impressive. After one round of difficulty climbing processing, the average difficulty score of the questions increased from 5.02 to 5.98 (out of 10). More importantly, the proportion of difficult questions scoring above 6.0 increased significantly from 47.2% to 76.8%. This change is not only reflected in the numbers but also in the actual application effects: AI models trained with difficulty-climbed questions performed significantly better in math reasoning tests than those trained with original questions.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">To better understand this process, we can look at a specific example. Suppose the original question is a relatively simple problem about trigonometric functions. The system may introduce an advanced theorem from complex analysis and incorporate a concept from combinatorial mathematics, ultimately generating a comprehensive problem that requires the use of trigonometric functions, complex analysis, and combinatorial thinking. Such a question not only maintains the mathematical spirit of the original question but also greatly increases the complexity and depth of the problem-solving process.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">III. Experimental Validation: Teaching Effectiveness Proven by Data\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">To verify the practical effect of the SAND-Math system, the research team conducted a series of comparative experiments, similar to conducting a teaching effectiveness comparison study. They selected the Qwen2.5-32B AI model as the \"student\" and trained it with different math question sets, observing which training method would allow this AI \"student\" to perform better in math exams.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The experiment design was very clever. The research team first established a benchmark test, using the LIMO dataset (a known high-quality math question bank) to train the AI model. This is like providing students with a classic math exercise book. Then, they trained the same AI model with questions generated by SAND-Math, other existing math question banks, and combinations of both, observing the impact of different training methods on the model's math abilities.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The evaluation criteria selected several representative math competitions: the 2024 and 2025 American Invitational Mathematics Examination (AIME), the American Mathematics Competitions (AMC), and the MATH-500 test set. These tests are like different types of math exams, capable of assessing the AI model's math reasoning abilities from multiple angles.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The experimental results were impressive. When using SAND-Math questions as supplementary training materials, the AI model's performance on the AIME25 test improved from 71.50% to 73.32%. More impressively, this improvement was greater than that achieved by using any other existing math question bank. Specifically, the improvement effect of SAND-Math was 17.85 percentage points higher than that of the second-best synthetic dataset, which is a significant improvement in the field of AI training.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The research team also specifically tested the independent effect of the difficulty climbing technique. They found that models trained with difficulty-climbed questions performed better than those trained with original questions. On the AIME25 test, the difficulty climbing technique improved the model's score from 46.38% to 49.23%. This result clearly shows that it's not about having more questions, but about having more challenging ones to achieve better training results.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The significance of these experimental results goes beyond numerical improvements.\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\">They prove an important educational philosophy: high-quality practice materials are more effective in improving learning outcomes than a large amount of repetitive practice.\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Just like the principle of \"progressive overload\" in physical training, only by constantly increasing the challenge can we truly enhance our abilities.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">IV. Technical Implementation Details: The Technical Secrets Behind Building a Math Question Factory\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The technical implementation of the SAND-Math system is like building a highly automated factory, where every link requires precise technical configuration and quality control mechanisms. The research team invested a lot of effort in technical choices and system design to ensure the entire process is efficient and reliable.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The core engine of the system chose the DeepSeek-R1 model as the main \"math question creator.\" This choice was not arbitrary but based on the model's outstanding performance in mathematical reasoning. During the question generation phase, the system set the creative parameters (temperature) of the model to 0.8, which is like adjusting the creativity knob, ensuring both the diversity of the generated content and avoiding overly random outputs. During the answer generation phase, the parameter was adjusted to 0.6, slightly reducing randomness to ensure the accuracy of the answers.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The quality control section used the Llama-3.3-70B-Instruct model as the \"quality inspector.\" This model is responsible for verifying the consistency of answers, identifying duplicate content, and evaluating the difficulty of questions. To improve the reliability of the assessment, each question's difficulty score needs to be evaluated three times independently, and then the average is taken. This approach is like having multiple experts evaluate scores independently, and then making a comprehensive judgment, greatly improving the objectivity of the assessment results.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The de-duplication and plagiarism detection section used advanced semantic hashing technology. The system uses the semhash framework combined with the minisilab\u002Fpotion-base-8M model to detect content duplication at a similarity threshold of 0.99. This process is like a library's plagiarism detection system, which can identify questions that have different expressions but the same essence. Interestingly, the system also has a web search function, which uses each generated question as a search keyword to find similar content on the internet, ensuring that the generated questions are indeed original.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The hardware configuration of the entire system is quite luxurious: 8 AMD InstinctTM MI300X GPUs form a computing cluster, and all models run on a single node, ensuring processing speed and stability. To optimize performance, the team also used the ZeRO-3 technology of the DeepSpeed framework for memory management. Although these technical details may seem abstract to ordinary users, they collectively ensure that the system can efficiently handle the generation and screening of a large number of math problems.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">It is worth mentioning that the research team also developed a specialized mathematical knowledge classification system. They divided mathematical knowledge into multiple branches, including number theory, algebra, geometry, combinatorics, and probability theory, with detailed lists of theorems and concepts under each branch. This knowledge system is like a huge mathematical encyclopedia, providing rich resources for the difficulty climbing technology. When the system needs to increase the difficulty of a question, it can intelligently select relevant theorems and concepts from this knowledge base to ensure that the new question is mathematically reasonable and meaningful.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">V. Practical Applications and Future Prospects: The Innovation Path of Mathematics Education\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">The emergence of the SAND-Math system brings new possibilities to mathematics education. In traditional teaching models, teachers often spend a lot of time collecting and organizing math problems, and are limited by the quantity and quality of existing question banks.\u003C\u002Fstrong>\u003Cstrong style=\"font-size: 18px;\">Now, with a system that can infinitely generate high-quality math problems, educators can customize practice materials of different difficulties and types according to the specific needs of students.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">This technology's most direct application scenario is personalized learning. Each student's math foundation and learning progress are different, and traditional \"one-size-fits-all\" teaching methods often fail to meet the needs of all students. The SAND-Math system can generate appropriate practice questions based on the student's current level. For students with weaker foundations, the system can generate more basic questions; for students who have extra capacity, the system can provide more challenging problems. This is like customizing a personalized math workbook for each student.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Another important application is the training of AI educational assistants. As artificial intelligence is increasingly applied in the field of education, we need smarter and more capable AI to assist in math teaching. The high-quality question bank generated by SAND-Math provides valuable resources for training these AI educational assistants. By training on these carefully designed questions, AI assistants can better understand mathematical concepts and provide more accurate problem-solving guidance.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">From a more macro perspective, this technology may also change the way math competitions and exams are formulated. Traditional formulation work mainly relies on the experience and inspiration of experts, which is not only time-consuming and labor-intensive but also prone to issues of monotonous style. The SAND-Math system can serve as a valuable assistant to examiners, quickly generating a large number of candidate questions, which are then selected and refined by experts. This human-computer collaboration model ensures the quality of the questions while greatly improving the efficiency of question formulation.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Of course, any technology has its limitations. The quality of the questions generated by the SAND-Math system largely depends on the capability of the training model. If the \"teacher\" model has knowledge gaps or biases, the generated questions may also have corresponding issues. Additionally, the system currently focuses mainly on the mathematical difficulty and novelty of the questions, and there is still room for improvement in considering the educational value and practical application context of the questions.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The research team also acknowledges that their experiments were mainly based on relatively small samples. Although the results are encouraging, to fully realize the system's potential, further validation on larger datasets is needed. They plan to expand the scale of experiments in future research and explore the possibility of applying this technology to other subject areas.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Looking ahead, SAND-Math represents not just a math question generation tool, but also an important milestone in the development of educational technology. It demonstrates how artificial intelligence can creatively solve the problem of scarce educational resources, providing each learner with a richer and more personalized learning experience. As the technology continues to improve and the application scenarios expand, we have reason to believe that such intelligent educational tools will play an increasingly important role in the future education system.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In short, the true value of the SAND-Math system lies not in how many math problems it can generate, but in providing new possibilities for educational equity and personalized learning. When every student can get practice materials suitable for their level, and when every teacher can easily access high-quality teaching resources, our math education will become more efficient and interesting. This technology developed by the AMD team is quietly changing the face of math learning, making this abstract and beautiful subject more accessible and approachable. For readers who wish to gain a deeper understanding of the technical details, the complete research paper and related datasets are publicly available online, providing a solid foundation for further research and applications.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Q&A\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">Q1: What exactly is the SAND-Math system? What problems can it solve?\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A: SAND-Math is an AI math question generation system developed by AMD, like an tireless math teacher who can create high-quality math questions from scratch. It mainly solves the problem of scarce high-quality questions in current math education, especially the difficulty of finding enough diverse and challenging questions for training AI math assistants. The system not only generates questions but also automatically adjusts difficulty, providing suitable practice materials for learners of different levels.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">Q2: How good is the quality of the math questions generated by this system? Will it make mistakes?\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A: The system has a strict quality control mechanism. Each question must go through multiple rounds of verification: first, let the AI solve the same question from different angles, and only those with completely consistent answers can pass; then check for duplication with existing questions; finally, use a high-level AI model to test the difficulty. Experiments show that AI models trained with SAND-Math questions scored 17.85 percentage points higher in math tests than other methods, proving the high quality of the questions.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px;\">Q3: Can ordinary teachers and students use this system now?\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">A: Currently, the system is mainly used for research and AI model training, not yet a product for general users. However, the research team has already published related datasets and technical details, laying the foundation for developing educational tools for educators. In the future, it is very likely that teaching assistance tools based on this technology will appear, allowing teachers to customize practice questions according to students' needs.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(187, 187, 187);\">【News Source】Tech Player \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Famd%E5%8F%91%E5%B8%83sand-math-%E7%94%A8ai%E9%80%A0%E9%A2%98%E7%A5%9E%E5%99%A8%E8%AE%A4%E6%95%B0%E5%AD%A6%E5%AD%A6%E4%B9%A0%E6%9B%B4%E6%9C%89%E8%B6%A3-%E8%A7%A3%E5%86%B3%E7%90%86%E5%B7%A5%E7%A7%91%E6%95%99%E8%82%B2%E8%B5%84%E6%BA%90%E7%A8%80%E7%BC%BA%E9%9A%BE%E9%A2%98\u002Far-AA1JUS3R?ocid=BingNewsLanding&cvid=482a3bbb03ef4b21cd85ccb8b6128d11&ei=28\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">http:\u002F\u002Fu5a.cn\u002FRQoEQ\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（This article is reprinted by the website to provide readers with more information, and the content does not constitute investment or consumption advice. If there are any doubts about the facts of the article, please verify with the relevant parties, and the views of the article are not the views of the website, for reference only.）\u003C\u002Fspan>\u003C\u002Fp>","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A2ac047c1-0ef6-4f96-8e65-34cc5b9263f9%3A0.wav?Expires=1774838503&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=s3V2pDPzj%2B5KjjzQCffjOmaj9b0%3D","2ac047c1-0ef6-4f96-8e65-34cc5b9263f9",17292758]