[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fRVPCrxpUJtVZXvV-oYNsaX0cwc_ygYqUN2Hv7l1a-KY":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":13},1575,"哈佛大学学者发明“思维压缩器”让AI推理速度飞跃5倍","\u003Cimg alt=\"\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2026\u002F03\u002Fhistory\u002Fa52079ca12b346849c28b126f1c4dc4d.png\" width=\"754\" height=\"null\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\">来自哈佛大学、加州理工学院MIT和Hippocratic AI的研究团队在2025年11月发表了一项突破性成果，论文题目为\"ORION: Teaching Language Models to Reason Efficiently in the Language of Thought\"。这项研究发表在顶级会议的评审中，感兴趣的读者可以通过论文编号arXiv:2511.22891v1查询完整内容。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">当前的大型语言模型就像一个思维过度活跃的学生，为了解决一道简单的数学题，它会在草稿纸上写满密密麻麻的推理过程，用上千个词来表达本来几句话就能说清楚的思路。比如回答\"一个农夫有3只鸡和2头牛，它们总共有多少条腿？\"这样的问题，现在的AI模型会絮絮叨叨写上300多个词，反复验算、自我纠错，就像一个缺乏自信的学生在考试时不断检查答案。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">这种\"话痨式\"推理带来了严重的问题。首先是速度慢得让人抓狂，用户等待AI回答一个简单问题可能需要好几秒钟，因为系统在后台生成了大量冗余的思考文字。其次是成本高昂，每多生成一个词就意味着更多的计算资源消耗，就像出租车按里程计费一样，绕远路的代价最终都要用户买单。更重要的是，这种冗长的推理过程往往充满矛盾和错误，就像一个人自言自语时越说越糊涂。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">研究团队受到了一个有趣的心理学理论启发——\"思维语言假说\"。这个理论认为，人类大脑进行复杂思考时并不是用我们平时说话的自然语言，而是用一种更加简洁、符号化的\"内在语言\"，就像程序员写代码时使用的简洁命令一样。当你心算\"8+5\"时，大脑并不会默念\"八加上五等于十三\"这样的完整句子，而是用某种更直接的符号运算。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">基于这个洞察，研究团队开发了一套名为\"Mentalese\"（心智语言）的压缩推理系统。这就像给AI安装了一个\"思维压缩器\"，让它学会用最简洁的符号来表达推理过程。原本需要300个词的农夫数腿问题，在新系统中只需要10个词就能搞定：\"鸡2腿，牛4腿...3×2+2×4=14\"。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">然而，仅仅让AI学会简洁表达还不够，研究团队面临一个关键挑战：如何在保持简洁的同时不损失推理的准确性？这就像要求一个健谈的人突然变得惜字如金，初期肯定会出现表达不清、遗漏要点的问题。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">为了解决这个难题，研究团队设计了一种巧妙的训练方法，叫做\"短优偏好优化\"（SLPO）。这种方法的核心思路类似于一个明智的老师评判学生作业：如果两份作业都答对了，那么表达更简洁的那份会得到更高的分数；但如果只有冗长的那份答对了，老师绝不会因为它太啰嗦就给低分。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">这种训练策略的精妙之处在于它的自适应性。面对简单问题时，系统会倾向于用最简洁的方式表达；但遇到复杂问题时，系统不会被强制压缩，而是允许使用更多的推理步骤。这就像一个经验丰富的医生，看感冒时几分钟就能下诊断，但面对疑难杂症时会进行详细的检查和分析。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">研究团队的训练过程分为两个阶段，就像培养一个高效思考者的完整流程。第一阶段是\"符号化学习\"，让AI系统熟悉这种简洁的表达方式，就像教孩子学会用数学符号而不是文字来表达数学概念。第二阶段是\"强化优化\"，通过奖励机制让AI在保持准确性的前提下追求表达的简洁性。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">实验结果令人印象深刻。在数学推理任务上，新开发的ORION模型实现了4到16倍的文字压缩比，推理速度提升了5倍，训练成本降低了7到9倍。更重要的是，准确率只下降了2到10个百分点，这在工程实践中是完全可以接受的代价。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">以AIME数学竞赛为例，原本的DeepSeek R1模型需要平均7481个词来解答一道题目，而ORION模型只需要184个词就能达到相近的正确率。这种压缩不是简单的删减，而是真正提取了推理的本质，去除了冗余和噪音。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">研究团队还与最先进的商业模型进行了对比测试。令人惊讶的是，他们只有15亿参数的ORION模型，在保持2倍压缩率的情况下，准确率竟然比GPT-4o和Claude-3.5这样的大型模型还要高出5%。这就像一台小巧的跑车在燃油经济性和速度上同时击败了大型SUV。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">这项技术的应用前景非常广阔。在实时对话系统中，用户不再需要等待AI冗长的\"思考\"过程，可以获得几乎即时的智能回复。在教育领域，AI辅导系统可以提供更加清晰、直接的解题思路，而不是让学生被冗长的推理过程搞得更加困惑。在专业咨询服务中，AI可以快速给出要点明确的建议，提高工作效率。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">不过，这项技术也存在一些局限性。研究团队发现，过度的压缩训练可能导致AI在面对真正复杂问题时推理深度不够。此外，当给AI设置了过长的生成限制时，它有时会\"退化\"回原来冗长的推理模式，就像一个已经养成简洁习惯的人在压力下又开始啰嗦。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">从技术发展的角度看，这项研究代表了AI领域一个重要的范式转变：从追求\"更多\"转向追求\"更好\"。过去几年，AI的发展主要依赖于增加模型规模和数据量，但这种方式面临着成本和效率的双重压力。ORION模型证明了，通过改进算法和训练方法，可以用更小的模型达到更好的效果。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">这种技术进步对整个AI产业的意义重大。它降低了部署智能系统的门槛，让更多中小企业能够承担AI应用的成本。同时，它也为AI在移动设备和边缘计算场景中的应用打开了新的可能性，因为压缩后的模型需要更少的计算资源和存储空间。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">研究团队还开源了包含4万个数学问题的\"MentaleseR-40k\"数据集，为其他研究者提供了宝贵的资源。这个数据集中的每个问题都被转换成了简洁的符号化推理形式，就像为AI研究社区提供了一本\"高效思维\"的教科书。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">展望未来，这项技术有可能催生新一代的AI助手，它们不仅更快、更省资源，还能提供更加清晰和直观的推理过程。这对于需要AI解释其决策过程的关键应用场景特别重要，比如医疗诊断、法律分析或金融投资建议。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">说到底，这项研究的最大价值在于它改变了我们对AI推理的理解。它证明了智能不在于话说得多，而在于思考得准。就像古人说的\"言简意赅\"，真正的智慧往往体现在用最少的话表达最深刻的思想。ORION模型让AI向着这个方向迈出了重要一步，为未来更加高效、实用的人工智能系统奠定了基础。这不仅是技术的进步，更是AI向人类思维模式学习的一次成功尝试。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Q&A\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Q1：Mentalese心智语言是什么？\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">A：Mentalese是研究团队开发的一种压缩推理语言，类似于数学公式那样用符号化的方式表达思维过程。它让AI不再用冗长的自然语言进行推理，而是用简洁的符号命令，比如用\"SET:w;EQ:abs(180-5.5*w)=110\"这样的格式来解决数学问题，从而大幅减少不必要的文字。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Q2：ORION模型的推理压缩技术会影响准确率吗？\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">A：会有一定影响，但在可接受范围内。ORION模型虽然将推理文字压缩了4到16倍，但准确率只下降了2到10个百分点。更重要的是，在某些测试中，ORION模型甚至比GPT-4o和Claude这样的大型模型准确率还要高出5%，证明简洁推理不等于推理能力降低。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Q3：短优偏好优化SLPO训练方法有什么特别之处？\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">A：SLPO的特别之处在于它不会盲目追求简洁而牺牲准确性。就像一个明智的老师，如果学生的两份作业都答对了，它会奖励更简洁的那份；但如果只有详细的作业答对了，它绝不会因为篇幅长而给低分。这种智能化的奖励机制让AI学会了在保证正确性的前提下追求表达效率。\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>\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(136, 136, 136);\">【新闻来源】MSN \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Far-AA1TIkbB?ocid=msedgntphdr&cvid=695e0341f673403199c8cca32fbc7b7b&cvpid=695e108483724f1b8b527a55c78319a5&ei\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(136, 136, 136);\"> https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Far-AA1RTDQy?ocid=BingHp01&cvid=6936317f054647a2afcd53fafcde084a&ei\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(136, 136, 136);\">（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>\u003Cp>\u003Cbr>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2026\u002F01\u002F0fba2f441b394facb535e00a74653958\u002FAI领域.jpg","https:\u002F\u002Fimage.51xinwei.com\u002F2026\u002F01\u002Fthumbs\u002F0fba2f441b394facb535e00a74653958\u002FAI领域.jpg",0,1,44,"2026-01-08 16:45",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%3A3a0ffa1a-7b26-4b59-89d6-fa112d4437b4%3A0.wav?Expires=1768846348&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=4Pcnc4r%2BM1cgZkeZSqgz5iH77EA%3D",15280832,"3a0ffa1a-7b26-4b59-89d6-fa112d4437b4","2026-01-08 16:43","Harvard University scholars have invented a \"thought compressor\" that makes AI reasoning 5 times faster","\u003Cimg alt=\"\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2026\u002F03\u002Fhistory\u002Fa52079ca12b346849c28b126f1c4dc4d.png\" width=\"754\" height=\"null\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\">A research team from Harvard University, California Institute of Technology, MIT and Hippocratic AI published a breakthrough in November 2025, with the paper titled \"ORION: Teaching Language Models to Reason Efficiently in the Language of Thought\". This study was reviewed at a top conference, and interested readers can query the full content using the paper number arXiv:2511.22891v1.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Current large language models are like overly active students, who write dense reasoning processes on scratch paper for solving simple math problems, using thousands of words to express ideas that could be explained in just a few sentences. For example, answering a question like \"A farmer has 3 chickens and 2 cows, how many legs do they have in total?\" modern AI models would write over 300 words, repeatedly checking and correcting themselves, like an insecure student constantly verifying answers during an exam.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">This \"talkative-style\" reasoning brings serious problems. First, it is extremely slow, making users wait several seconds for AI to answer a simple question because the system generates a lot of redundant thinking text in the background. Second, it is costly, as each additional word means more computational resources consumption, similar to a taxi charging by distance, where the cost of taking a detour ultimately has to be borne by the user. More importantly, this lengthy reasoning process often contains contradictions and errors, like a person talking to themselves getting more confused the longer they speak.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">The research team was inspired by an interesting psychological theory - the \"language of thought hypothesis.\" This theory suggests that when humans think about complex issues, they do not use natural language we usually speak, but rather a more concise and symbolic \"inner language,\" similar to the concise commands used by programmers when writing code. When you mentally calculate \"8+5,\" your brain does not silently recite \"eight plus five equals thirteen,\" but instead uses some more direct symbolic operations.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Based on this insight, the research team developed a compression reasoning system called \"Mentalese\" (mental language). It's like installing a \"thought compressor\" on AI, enabling it to express reasoning processes using the most concise symbols. The problem of counting legs for a farmer, which originally required 300 words, can now be solved in just 10 words in the new system: \"chickens 2 legs, cows 4 legs... 3×2 + 2×4 = 14.\"\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">However, merely teaching AI to express concisely is not enough. The research team faced a key challenge: how to maintain reasoning accuracy while keeping it concise? This is like asking a talkative person to suddenly become very succinct, which initially will certainly lead to unclear expressions and missing key points.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">To solve this problem, the research team designed a clever training method called \"Short and Optimal Preference Optimization\" (SLPO). The core idea of this method is similar to a wise teacher grading student assignments: if two assignments are both correct, the one with a more concise expression will receive a higher score; but if only the verbose one is correct, the teacher will not give it a low score just because it's too wordy.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">The elegance of this training strategy lies in its adaptability. For simple questions, the system tends to express them in the most concise way; but for complex questions, the system is not forced to compress, but is allowed to use more reasoning steps. This is like an experienced doctor who can diagnose a common cold in minutes, but for complicated cases, will conduct detailed examinations and analyses.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">The research team's training process consists of two stages, much like the complete process of cultivating an efficient thinker. The first stage is \"symbolic learning,\" allowing the AI system to get familiar with this concise expression style, like teaching a child to use mathematical symbols instead of words to express mathematical concepts. The second stage is \"reinforcement optimization,\" where the AI pursues conciseness in expression while maintaining accuracy through a reward mechanism.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">The experimental results are impressive. In mathematical reasoning tasks, the newly developed ORION model achieved a 4 to 16 times text compression ratio, with a 5 times increase in reasoning speed and a 7 to 9 times reduction in training costs. More importantly, the accuracy rate only dropped by 2 to 10 percentage points, which is completely acceptable in engineering practice.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Taking the AIME mathematics competition as an example, the original DeepSeek R1 model needed an average of 7481 words to answer a question, while the ORION model only needed 184 words to achieve a similar accuracy rate. This compression is not a simple deletion, but truly extracts the essence of reasoning, removing redundancy and noise.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">The research team also conducted comparative tests with the most advanced commercial models. Surprisingly, their ORION model with only 1.5 billion parameters achieved an accuracy rate 5% higher than large models like GPT-4o and Claude-3.5 while maintaining a 2 times compression rate. This is like a compact sports car defeating a large SUV in both fuel economy and speed.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">The application prospects of this technology are very broad. In real-time conversation systems, users no longer need to wait for AI's long \"thinking\" process and can get almost instant intelligent responses. In the education sector, AI tutoring systems can provide clearer and more direct problem-solving approaches, rather than confusing students with long-winded reasoning processes. In professional consulting services, AI can quickly offer clear-cut advice, improving work efficiency.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">However, this technology also has some limitations. The research team found that excessive compression training may lead to insufficient reasoning depth when AI faces truly complex problems. Additionally, when AI is given excessively long generation limits, it sometimes \"degenerates\" back into its original verbose reasoning mode, similar to someone who has developed a habit of being concise starting to ramble again under pressure.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">From a technological development perspective, this research represents an important paradigm shift in the AI field: from pursuing \"more\" to pursuing \"better.\" In recent years, AI development has mainly relied on increasing model size and data volume, but this approach faces dual pressures of cost and efficiency. The ORION model proves that by improving algorithms and training methods, better results can be achieved with smaller models.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">This technological advancement holds significant meaning for the entire AI industry. It lowers the threshold for deploying intelligent systems, enabling more small and medium-sized enterprises to afford AI applications. At the same time, it opens up new possibilities for AI applications on mobile devices and edge computing scenarios, as compressed models require fewer computational resources and storage space.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">The research team also open-sourced the \"MentaleseR-40k\" dataset containing 40,000 math problems, providing valuable resources for other researchers. Each problem in this dataset is converted into a concise symbolic reasoning form, like providing an \"efficient thinking\" textbook for the AI research community.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Looking ahead, this technology could give rise to a new generation of AI assistants that are not only faster and more resource-efficient but also provide clearer and more intuitive reasoning processes. This is especially important for critical application scenarios where AI needs to explain its decision-making process, such as medical diagnosis, legal analysis, or financial investment advice.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">In the end, the greatest value of this research lies in changing our understanding of AI reasoning. It proves that intelligence is not about speaking a lot, but about thinking accurately. Like the ancient saying \"speak briefly but convey the meaning,\" true wisdom often lies in expressing profound thoughts with the fewest words. The ORION model has taken an important step in this direction, laying the foundation for future more efficient and practical AI systems. This is not just a technological advancement, but also a successful attempt for AI to learn from human thinking patterns.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Q&A\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Q1: What is Mentalese mental language?\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">A: Mentalese is a compressed reasoning language developed by the research team, similar to mathematical formulas that express thinking processes in symbolic form. It enables AI to no longer reason in verbose natural language, but instead use concise symbolic commands, such as the format \"SET:w;EQ:abs(180-5.5*w)=110\" to solve math problems, thus significantly reducing unnecessary text.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Q2: Will ORION model's reasoning compression technology affect accuracy?\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">A: It will have some impact, but within an acceptable range. Although the ORION model compresses reasoning text by 4 to 16 times, the accuracy rate only drops by 2 to 10 percentage points. More importantly, in some tests, the ORION model even achieved a 5% higher accuracy rate than large models like GPT-4o and Claude, proving that concise reasoning does not mean reduced reasoning ability.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">Q3: What is special about the SLPO training method of short and optimal preference optimization?\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\">A: The special aspect of SLPO is that it does not blindly pursue conciseness at the expense of accuracy. Just like a wise teacher, if two students' assignments are both correct, it rewards the more concise one; but if only the detailed assignment is correct, it will not give it a low score just because it is lengthy. This intelligent reward mechanism allows AI to learn to pursue expression efficiency while ensuring correctness.\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>\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(136, 136, 136);\">【News Source】MSN \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Far-AA1TIkbB?ocid=msedgntphdr&cvid=695e0341f673403199c8cca32fbc7b7b&cvpid=695e108483724f1b8b527a55c78319a5&ei\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(136, 136, 136);\"> https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Far-AA1RTDQy?ocid=BingHp01&cvid=6936317f054647a2afcd53fafcde084a&ei\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(136, 136, 136);\">（This article is reprinted by the website to provide readers with more information and news. The content involved does not constitute investment or consumer suggestions. If there are any doubts about the facts of the article, please verify with the relevant parties. The views of the article are not the views of the website, and are for reference only.）\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A91c0ae58-8927-4c85-83a1-85eafb575df1%3A0.wav?Expires=1774838432&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=3SMKuL9ROGfXVNPs4V4kT2nIYwE%3D","91c0ae58-8927-4c85-83a1-85eafb575df1"]