[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fXDbWQJoMmtrFtvBG4ZP6aCNrT-g2EApUTKQzJVx_bhE":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},1418,"三星AI研究院发布开源TRM模型：小参数大作为，结构化推理领域挑战顶尖大模型","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">三星高级人工智能研究院近日公布了一项突破性成果——一款名为微型递归模型（TRM）的开源AI系统，其参数规模仅700万，却在特定结构化推理任务中展现出与谷歌Gemini 2.5 Pro等万倍参数量级模型相抗衡的实力。这一发现为AI领域“小而精”模型的发展提供了全新思路。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">该模型由研究院资深研究员Alexia Jolicoeur-Martineau团队开发，其核心设计理念颠覆了传统AI架构。不同于依赖多层网络协作的分层推理模型（HRM），TRM采用仅含两层的极简结构，通过“递归推理”机制实现性能突破。模型会反复检验自身输出的预测结果，在每轮迭代中修正前序错误，直至获得稳定解。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这种设计使TRM在保持低计算资源消耗的同时，通过深度迭代模拟了大型模型的复杂推理过程。研究团队形象地将其策略概括为“以递归替代规模”，即通过算法优化而非参数堆砌实现性能提升。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在实测环节，TRM在多个结构化任务基准测试中表现亮眼：极限数独测试准确率达87.4%，困难迷宫任务准确率85%，抽象推理能力测试ARC-AGI中取得45%准确率，ARC-AGI-2测试准确率则为8%。这些数据表明，其性能已接近或超越DeepSeek R1、o3-mini等知名大模型，尽管参数规模不足后者的0.01%。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">但研究团队明确指出，TRM的卓越表现具有特定适用范围。该模型专为数独、迷宫等规则明确的网格类问题优化，在需要开放式语言生成的场景中并不适用。其优势在于封闭环境下的精确逻辑推理，而非通用语言处理能力。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">深入分析显示，TRM的成功源于对“极简主义”的极致追求。实验表明，增加模型层数或参数规模反而会导致小数据集上的过拟合现象，降低实际性能。双层结构与递归机制的组合，恰好实现了复杂度与效率的最优平衡。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">目前，TRM的全部代码、训练脚本及测试数据集已通过MIT许可证在GitHub平台开源。全球开发者可自由获取、修改并用于商业项目，这一举措预计将推动结构化推理AI在工业、教育等领域的快速应用。\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>\u003Cspan style=\"color: rgb(187, 187, 187);\">【新闻来源】IT之家 \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Far-AA1OcE0J?ocid=BingNewsSerp\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Far-AA1OcE0J?ocid=BingNewsSerp\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>\u003Cp>\u003Cbr>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F10\u002F83fb4ff1d53e4051911fe385db6df2d7\u002FAI领域.jpg","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F10\u002Fthumbs\u002F83fb4ff1d53e4051911fe385db6df2d7\u002FAI领域.jpg",0,1,55,"2025-10-13 22:05",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%3A3b6c0618-24c3-4154-9033-b770acbafae6%3A0.wav?Expires=1761209237&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=JpjXOqsBKC6b6yyAmeoK29oaGoI%3D",4731782,"3b6c0618-24c3-4154-9033-b770acbafae6","2025-10-13 22:03","Samsung AI Institute releases open-source TRM model: small parameters, great performance, challenging top large models in structured reasoning field","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">Samsung Advanced AI Institute recently announced a breakthrough - an open-source AI system called the Tiny Recursive Model (TRM), which has only 7 million parameters but demonstrates the ability to compete with models such as Google Gemini 2.5 Pro, which have a thousand times more parameters, in specific structured reasoning tasks. This discovery provides a new idea for the development of \"small but precise\" models in the AI field.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The model was developed by the team of senior researcher Alexia Jolicoeur-Martineau from the institute. Its core design concept subverts traditional AI architecture. Unlike hierarchical reasoning models (HRM) that rely on multi-layer network collaboration, TRM adopts an extremely simplified structure with only two layers and achieves performance breakthroughs through a \"recursive reasoning\" mechanism. The model repeatedly checks its own output predictions and corrects previous errors in each iteration until it obtains a stable solution.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">This design allows TRM to maintain low computational resource consumption while simulating the complex reasoning process of large models through deep iteration. The research team vividly summarizes its strategy as \"replacing scale with recursion,\" achieving performance improvement through algorithm optimization rather than parameter stacking.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In practical testing, TRM performed well in multiple structured task benchmark tests: it achieved an accuracy rate of 87.4% in the extreme Sudoku test, 85% in the difficult maze task, and 45% in the abstract reasoning ability test ARC-AGI, while achieving 8% accuracy in the ARC-AGI-2 test. These data indicate that its performance is close to or exceeds well-known large models such as DeepSeek R1 and o3-mini, despite having less than 0.01% of their parameter scale.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">However, the research team clearly pointed out that TRM's outstanding performance has a specific application scope. The model is optimized for grid-like problems with clear rules such as Sudoku and mazes, and is not suitable for scenarios requiring open-ended language generation. Its advantage lies in precise logical reasoning in closed environments, rather than general language processing capabilities.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In-depth analysis shows that TRM's success stems from an extreme pursuit of minimalism. Experiments show that increasing the number of model layers or parameter scale can lead to overfitting on small datasets, reducing actual performance. The combination of a two-layer structure and recursive mechanism achieves an optimal balance between complexity and efficiency.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Currently, all the code, training scripts, and test dataset of TRM are open-sourced on the GitHub platform under the MIT license. Global developers can freely obtain, modify, and use them for commercial projects. This move is expected to promote the rapid application of structured reasoning AI in industries, education, and other fields.\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>\u003Cspan style=\"color: rgb(187, 187, 187);\">【News source】IT Home \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Far-AA1OcE0J?ocid=BingNewsSerp\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002Far-AA1OcE0J?ocid=BingNewsSerp\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（This article is forwarded by this site to provide readers with more information and news. The content involved does not constitute investment or consumer advice. If there are any questions about the facts of the article, please verify with the relevant parties. The views of the article are not the views of this site, 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%3A51bc8329-2269-47bc-9eed-5172f1c1a9f4%3A0.wav?Expires=1774838460&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=oe0TKlW3NNQz1KqtAVJ27DNDimQ%3D","51bc8329-2269-47bc-9eed-5172f1c1a9f4",6094988]