谷歌大模型获重大突破，首次拿下国际奥数银牌
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-axegupay5k/42297e2e74b3449d9e05f6898958ff33~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1722928825&x-signature=dx4G1X3nlnx9DkoUq48yBBoxx5M%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">谷歌DeepMind<span style="color: black;">颁布</span>了一项重大成绩，<span style="color: black;">运用</span>AlphaProof和AlphaGeometry 2两个混合大模型参加了2024年国际数学奥林匹克竞赛（IMO）并<span style="color: black;">得到</span>了银牌。</span></span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-6w9my0ksvp/7ddbda4a36364819823acf6454b539b1~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1722928825&x-signature=PaLIPif8fkeAt%2BOtapR4bFWCZ3g%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">IMO是最古老、权威的数学竞赛，每年都会有来自世界各地精英级数学家参与，<span style="color: black;">同期</span><span style="color: black;">亦</span>是AI模型的竞技场，是衡量其数学推理能力的最佳平台。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">在今年的比赛中，一共有6道数学题，每答对一道获7分。AlphaProof<span style="color: black;">处理</span>了两道代数和一个数论问题；AlphaGeometry 2答对了一道几何题，一共<span style="color: black;">得到</span>了28分仅比金牌少1分。</span></span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-6w9my0ksvp/6420b8ff23b041f8a9060f544715d3a3~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1722928825&x-signature=4i9JXpL26HyUd2TLb40XxMGF174%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">值得一提的是，AlphaProof解答了今年IMO比赛最难的一道题，609位参赛者<span style="color: black;">仅有</span>5个人给出了正确答案。</span></span></p>
<h1 style="color: black; text-align: left; margin-bottom: 10px;">AlphaGeometry 2</h1>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">早在今年1月17日谷歌便发布了AlphaGeometry模型，并在30道几何奥林匹克测试题中答对了25道，</span><strong style="color: blue;"><span style="color: black;">这比之前由中国著名数学家、计算机家-吴文俊提出的最先进<span style="color: black;">办法</span>还多15道，仅比人类金牌得主少0.9分</span></strong><span style="color: black;">。</span></span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-6w9my0ksvp/a1d3b99946d041c7b7729de32750697c~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1722928825&x-signature=vS04zYDdp56RnYa4Qstg6CQ4TPE%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">AlphaGeometry的核心在于其神经符号框架，一个能够自动<span style="color: black;">处理</span>欧几里得平面几何问题的<span style="color: black;">繁杂</span>模型，</span><strong style="color: blue;"><span style="color: black;">绕过了传统<span style="color: black;">设备</span>学习<span style="color: black;">办法</span>中对<span style="color: black;">海量</span>人类证明数据的依赖，实现了从零<span style="color: black;">起始</span>的自我学习，使得AlphaGeometry能够生成<span style="color: black;">海量</span>的合成定理和证明</span></strong><span style="color: black;">，构建出一个有向无环图，<span style="color: black;">暗示</span>所有<span style="color: black;">达到</span>到的结论。</span></span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-6w9my0ksvp/074daabe4b4a430093493f1d8d8e3f4f~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1722928825&x-signature=klafNko0b6WUPoRxSQxmjGWHKYQ%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;"><span style="color: black;">证明搜索是AlphaGeometry神经符号框架的核心之一，这是一个循环过程，语言模型和符号推理引擎交替运行</span></strong><span style="color: black;">。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">语言模型<span style="color: black;">按照</span>问题<span style="color: black;">描述</span>生成新的辅助构造，符号推理引擎则利用这些构造扩展其推理闭包，直到达到结论或达到最大迭代次数。在每一次循环中，语言模型都会<span style="color: black;">按照</span>当前的证明状态和已有的构造生成一个新的辅助构造。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"><span style="color: black;">而后</span>，符号推理引擎会将这个新构造纳入<span style="color: black;">思虑</span>，尝试<span style="color: black;">经过</span><span style="color: black;">规律</span>推理来接近或达到结论。<span style="color: black;">倘若</span>推理引擎能够证明定理，搜索过程结束；<span style="color: black;">倘若</span>不行，循环将继续，语言模型将生成另一个辅助构造。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;"><span style="color: black;">证明修剪是AlphaGeometry<span style="color: black;">另一</span>一个重要功能</span></strong><span style="color: black;">。在自动证明过程中，可能会生成<span style="color: black;">有些</span>不必要的辅助构造，这些构造虽然不是错误的，但它们可能会使证明过程变得冗长和<span style="color: black;">繁杂</span>。<span style="color: black;">经过</span>证明修剪<span style="color: black;">能够</span>去除这些不必要的构造，确<span style="color: black;">保准</span>明的简洁性和可读性。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">AlphaGeometry<span style="color: black;">经过</span>穷举<span style="color: black;">实验</span>和错误的<span style="color: black;">办法</span>进行证明修剪。模型会尝试丢弃<span style="color: black;">每一个</span>辅助点的子集，并重新运行符号推理引擎，以验证在<span style="color: black;">无</span>这些辅助点的<span style="color: black;">状况</span>下<span style="color: black;">是不是</span>仍然能够达到结论。<span style="color: black;">经过</span>这种方式，模型能够找到并返回所有可能证明中的最短路径。</span></span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-6w9my0ksvp/d129710e93c348fdb90f0d0ad6788eca~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1722928825&x-signature=FRveclpShKXXwV4lmqRQ9P8m%2B8k%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;"><span style="color: black;">另外</span>，在合成证明生成中，AlphaGeometry引入了“依赖差异”的概念，这一概念<span style="color: black;">准许</span>系统生成构建辅助点的证明<span style="color: black;">过程</span>，并超越了纯符号推理的范围。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">辅助构造是几何证明中的一个关键挑战，它<span style="color: black;">表率</span>了证明过程中的无限分支<span style="color: black;">原因</span>。<span style="color: black;">经过</span>这种方式，AlphaGeometry能够生成几乎无限的证明变体，为深度学习模型<span style="color: black;">供给</span>了丰富的训练数据。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">而AlphaGeometry 2是在一代的<span style="color: black;">基本</span>之上进行了<span style="color: black;">海量</span>迭代和技术创新，<span style="color: black;">运用</span>了谷歌自研的Gemini<span style="color: black;">做为</span>语言模型，并在比一代多一个数量级的合成数据上从头<span style="color: black;">起始</span>训练。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">AlphaGeometry 2<span style="color: black;">运用</span>的符号引擎比其前身快两个数量级。当<span style="color: black;">显现</span>新问题时，<span style="color: black;">运用</span>一种新的知识共享机制来实现<span style="color: black;">区别</span>搜索树的高级组合，以<span style="color: black;">处理</span>更<span style="color: black;">繁杂</span>的数学<span style="color: black;">困难</span>。</span></span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-6w9my0ksvp/6689fc7483844221b0454edebc139f2a~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1722928825&x-signature=OjRQf2yB38V3mhojmNgbx7mi7jc%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">在今年IMO比赛之前，AlphaGeometry 2<span style="color: black;">能够</span><span style="color: black;">处理</span>过去25年所有IMO几何问题的83%，而一代<span style="color: black;">处理</span>率<span style="color: black;">仅有</span>53%。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">在今年正式比赛中，</span><strong style="color: blue;"><span style="color: black;">AlphaGeometry 2在收到几何问题后，仅用19秒便<span style="color: black;">处理</span>了这道<span style="color: black;">困难</span></span></strong><span style="color: black;">，达到了人类难以企及的超<span style="color: black;">有效</span>率。</span></span></p>
<h1 style="color: black; text-align: left; margin-bottom: 10px;">AlphaProof</h1>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">AlphaProof是谷歌最新<span style="color: black;">研发</span>的一个专门用于形式数学推理的模型，其核心特点是结合了预训练语言模型和AlphaZero强化学习算法，能够在<span style="color: black;">繁杂</span>的数学问题上展现出强大的推理能力。</span></span></p>
<div style="color: black; text-align: left; margin-bottom: 10px;"><img src="https://p3-sign.toutiaoimg.com/tos-cn-i-6w9my0ksvp/30fa8ec3fd2246b0be0ddc375e0d805f~noop.image?_iz=58558&from=article.pc_detail&lk3s=953192f4&x-expires=1722928825&x-signature=jp2HemtCRumexSEDI3MLiwpkOoQ%3D" style="width: 50%; margin-bottom: 20px;"></div>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">AlphaProof的工作原理是<span style="color: black;">运用</span>形式语言Lean来进行数学证明。形式语言的<span style="color: black;">优良</span>在于<span style="color: black;">能够</span>严格验证推理<span style="color: black;">过程</span>的正确性，但传统上受限于人工编写的数据量较少。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">为<span style="color: black;">认识</span>决这一<span style="color: black;">困难</span>，</span><strong style="color: blue;"><span style="color: black;">AlphaProof<span style="color: black;">运用</span>了一个经过微调的Gemini大模型，将自然语言问题自动转换为形式语言表述，从而创建了一个<span style="color: black;">包括</span><span style="color: black;">各样</span>难度和数学主题的大规模形式问题库</span></strong><span style="color: black;">。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">在面对一个新问题时，AlphaProof会生成<span style="color: black;">处理</span><span style="color: black;">方法</span>候选，<span style="color: black;">而后</span><span style="color: black;">经过</span>在Lean中搜索可能的证明<span style="color: black;">过程</span>来证明或反驳这些候选解。<span style="color: black;">每一个</span>被<span style="color: black;">发掘</span>和验证的证明都会用来强化AlphaProof的语言模型，<span style="color: black;">加强</span>它<span style="color: black;">处理</span>后续更具挑战性问题的能力。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><span style="color: black;">为了备战2024年IMO，AlphaProof在比赛前的几周内进行了密集数据训练，证明或反驳了数百万个问题，涵盖了广泛的难度和数学主题<span style="color: black;">行业</span>。</span></span></p>
<p style="font-size: 16px; color: black; line-height: 40px; text-align: left; margin-bottom: 15px;"><span style="color: black;"><strong style="color: blue;"><span style="color: black;">谷歌<span style="color: black;">暗示</span></span></strong><span style="color: black;">，<span style="color: black;">日前</span>的AI模型在<span style="color: black;">处理</span><span style="color: black;">通常</span>的数学问题时仍然存在困难，<span style="color: black;">重点</span>是<span style="color: black;">由于</span>推理的局限性和训练数据不足。<span style="color: black;">这次</span>参数的两个混合大模型<span style="color: black;">已然</span>具备数学推理的AGI（通用人工智能）能力，<span style="color: black;">能够</span><span style="color: black;">帮忙</span>数学专家<span style="color: black;">发掘</span>新的解题<span style="color: black;">办法</span>，<span style="color: black;">同期</span><span style="color: black;">亦</span>是AI在数学<span style="color: black;">行业</span>的重大技术突破。</span></span></p>
我们有着相似的经历，你的感受我深有体会。
页:
[1]