<p>，/PRNewswire/ -- 随着企业数据资产管理方式从简单的积累转向价值提取，人工智能的角色也相应发生了变化：它不再仅限于基本的数据处理和分析，而是通过基于企业真实商业逻辑的稳定、可靠和确定性的推理能力来定义。随着大型模型深入集成到行业应用中，出现了一个挑战：如何超越概率生成，使人工智能能够在复杂决策和现实商业流程中协助可靠的知识检索和操作计划。</p>
<p>用友 AI 实验室发布了大型本体模型（LOM）。LOM 基于构建 - 对齐 - 推理（CAR）的集成端到端架构，使人工智能首次能够自主地从原始企业数据构建结构化的商业逻辑系统，并在此系统上进行高精度推理。对真实企业数据集的实验结果证明了这种方法的有效性：LOM-4B 在本体补全任务中达到了 88.8% 的准确率，在复杂图推理任务中达到了 94% 的准确率，显著优于现有主流大型语言模型。这些结果指出了一种可扩展和确定性部署企业级人工智能的新范式。</p>
<p><strong>超越参数竞赛：赋能人工智能构建企业 “商业逻辑宇宙”</strong></p>
<p>多年来，在企业场景中采用大型模型在很大程度上遵循了 “更多参数意味着更好性能” 的线性假设。然而，在实际部署中，这种方法往往不尽如人意，例如产生不稳定的推理、不一致的输出，以及与商业逻辑的持续脱节。核心原因在于传统大型模型依赖于概率性标记预测，缺乏对企业系统的结构化理解，甚至缺乏自主构建符合企业实际业务需求的逻辑框架的能力。</p>
<p>用友 LOM 的核心突破在于超越单纯的参数扩展，而是赋能人工智能构建自主的逻辑系统。就像一位资深领域专家，LOM 能够从零散的结构化和非结构化数据中组织商业实体、属性及其相互关系，形成一个连贯的企业本体。该本体作为一个与企业所有推理任务的真实操作相一致的 “商业逻辑宇宙”，通过一个结构化、内部一致的环境展开所有推理，从根本上确保推理的确定性。</p>
<p>在本体构建阶段，LOM 同时处理存储在数据库中的结构化数据和非结构化文本文件。通过多阶段的生成和验证流程，它将分散的商业信息转换为标准化的、机器可解释的本体结构，同时通过迭代验证确保逻辑一致性。例如，从人力资源中的组织层级，到财务账户的关联，再到供应链中的上下游依赖关系，LOM 系统地重构了嵌入企业数据中的基础逻辑，以支持下一步操作的明确和可执行性。</p>
<p><strong>语义与结构深度融合的三方集成架构</strong></p>
<p>虽然自主本体构建构成了 LOM 的基础，但语义与结构的精确对齐使得这个 “逻辑宇宙” 真正反映了企业业务的现实。企业知识在语义多样性和结构复杂性方面具有特征。传统模型通常将文本理解和结构推理分开处理，导致信息丢失和推理不准确。</p>
<p>通过构建 - 对齐 - 推理的三方集成 CAR 架构，LOM 动态地将这三个过程整合为一个统一的认知框架。在对齐阶段，模型利用图感知编码器和强化学习，准确地将语义生成的信息与构建的本体结构匹配。这使得抽象图节点与企业的真实商业实体之间实现实时映射，同时识别本体的动态更新。当与用户的交互中出现新的商业知识和见解时，LOM 相应地调整本体结构，确保这个 “逻辑宇宙” 始终与企业业务的发展保持一致。</p>
<p>在推理阶段，LOM 摒弃了概率性猜测，而是使用自主构建的本体作为不可变的商业规则，在系统内执行严格的确定性推理。无论是执行最短路径和最小生成树等复杂图算法，还是导航多跳商业关系，LOM 都能产生可验证和可靠的结果。这种逻辑推理执行将人工智能从模拟推理转变为实际商业操作，完美适应需要极高准确性的企业场景，如金融、供应链和生产。</p>
<p><strong>在真实世界测试中超越主流模型：更少参数实现高认知密度</strong></p>
<p>范式创新带来的真正技术价值最终在真实场景中得到了验证。为此，用友 AI 实验室对 LOM 进行了全面测试，使用来自人力资源、金融、资产管理、制造和供应链等多个企业领域的生产数据。基准数据集包括 19 个图推理任务，将 LOM 的性能与主流大型语言模型进行了比较。</p>
<p>测试结果显示，LOM-4B 仅有 40 亿个参数，在所有任务上的平均准确率达到了 93%。而 LOM-32B 则拥有 320 亿个参数，进一步将准确率提高至 94%，特别是在需要确定性推理的任务中表现出色，如最短路径、循环检测和最小生成树。相比之下，主流模型虽然参数数量大幅增加，但主要在浅层语义任务中表现出色，这些任务只需表面模式和关联即可。然而，当面临复杂的结构推理挑战时，它们却显得力不从心，因为这些挑战中确定性逻辑和商业实体之间复杂关系深深交织在一起。在这些任务中，模型的表现急剧下降，部分任务的准确率接近零，突显了仅依赖概率模型在需要精确、基于规则的推理和结构一致性的任务中的固有限制。</p>
<p>这一结果强调了一个核心见解：企业级人工智能的真正力量不在于参数的积累，而在于认知密度。通过融合神经符号融合的技术理念，LOM 以更少的参数提供了增强的逻辑推理能力，证明了对于工业人工智能而言，逻辑密度比参数规模更具价值。对于企业而言，这意味着可以以更稳定可靠的方式获取与业务需求更为契合的人工智能推理能力，而无需在部署超大参数模型上投入巨额成本。</p>
<p><strong>7D 逻辑自主性：释放下一代企业人工智能的潜力</strong></p>
<p>在这一突破的基础上，用友 AI 实验室还开发了一个 10 维认知框架，概述了人工智能模型的演变。根据该框架，目前主流的大型模型和代理仍处于 6D 阶段，能够优化任务执行路径，但缺乏自主构建基础逻辑系统的能力。这一限制是企业应用中推理挑战的主要原因。</p>
<p>LOM 所实现的 7D 逻辑自主性标志着企业人工智能向更高维度演进的重要一步。它使人工智能能够从零开始构建推理的逻辑框架，实质上赋予其设定游戏规则的能力，而不仅仅是参与游戏。这一演变将人工智能从数据处理工具转变为真正的知识专家和企业业务逻辑的决策助手，为未来能够自主决策并主动触发业务执行的人工智能范式铺平了道路。</p>
<p>如今，随着企业数字化转型进入更高级的阶段，数据资产价值的全面实现越来越依赖于生成有意义的本体基础、采用准确的子图检索工具，从而为行动提供深入可靠的洞察。YonLOM 不仅提供确定性推理能力，还为将大型模型整合到行业实践中引入了一种新颖的解决方案。</p>
<p>展望未来，LOM 的 7D 逻辑自主性为进一步的进展奠定了基础，用友 AI 实验室将继续将模型从推理引擎演变为一个动态智能系统，能够提供更可靠的行动执行建议和战略决策。</p>
<p>欲获取更详细的见解，欢迎访问完整的预印本： https://chinaxiv.org/abs/202603.00072</p>
<p>来源：用友</p>
<p><img src="https://imageproxy.pbkrs.com/https://rt.prnewswire.com/rt.gif/query-RGF0ZUlkPTIwMjYwNDEwJk5ld3NJdGVtSWQ9Q04zMTQ1MyZUcmFuc21pc3Npb25fSWQ9MjAyNjA0MTAwMjQ4UFJfTkVXU19VU1BSX19fX19DTjMxNDUz?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://rt.prnewswire.com/rt.gif/query-RGF0ZUlkPTIwMjYwNDEwJk5ld3NJdGVtSWQ9Q04zMTQ1MyZUcmFuc21pc3Npb25fSWQ9MjAyNjA0MTAwMjQ4UFJfTkVXU19VU1BSX19fX19DTjMxNDUz"/></p>

用友网络

<p>用友网络人工智能实验室推出了大型本体模型（LOM），该模型使人工智能能够从原始企业数据中自主构建结构化的业务逻辑系统。该模型解决了传统大型模型的局限性，提供了稳定、可靠和确定性的推理。LOM 在本体补全任务中达到了 88.8% 的准确率，在复杂图推理任务中达到了 94% 的准确率，超越了现有模型。其三元架构整合了语义和结构，允许实时更新和精确推理，使其适用于金融、供应链和生产等高准确度的企业应用</p>

<p>, /PRNewswire/ -- As the way of managing enterprise data assets evolves from simple accumulation to value extraction, the role of AI has shifted accordingly: it is no longer limited to basic data processing and analysis, but it is now defined by its ability to perform stable, reliable, and deterministic reasoning grounded in real business logic of enterprises. As large models become deeply integrated into industry applications, a challenge has emerged: how to move beyond probabilistic generation and enable AI to assist with reliable knowledge retrieval and operation plan within complex decision making and real-world business processes.</p>
<div class="lb-trans"><p>，/PRNewswire/ -- 随着企业数据资产管理方式从简单的积累转向价值提取，人工智能的角色也相应发生了变化：它不再仅限于基本的数据处理和分析，而是通过基于企业真实商业逻辑的稳定、可靠和确定性的推理能力来定义。随着大型模型深入集成到行业应用中，出现了一个挑战：如何超越概率生成，使人工智能能够在复杂决策和现实商业流程中协助可靠的知识检索和操作计划。</p>
</div><p>Yonyou AI Lab has released the Large Ontology Model (LOM). Built on an integrated end-to-end architecture of Construct-Align-Reason (CAR), LOM enables AI, for the first time, to autonomously construct structured business logic system from raw enterprise data and perform high-precision reasoning on top of this system. Experimental results on real-world enterprise datasets demonstrate the effectiveness of this approach: LOM-4B achieved an accuracy of 88.8% in ontology completion tasks and 94% in complex graph reasoning tasks, significantly outperforming existing mainstream large language models. These results point out a new paradigm for the scalable and deterministic deployment of enterprise-grade AI.</p>
<div class="lb-trans"><p>用友 AI 实验室发布了大型本体模型（LOM）。LOM 基于构建 - 对齐 - 推理（CAR）的集成端到端架构，使人工智能首次能够自主地从原始企业数据构建结构化的商业逻辑系统，并在此系统上进行高精度推理。对真实企业数据集的实验结果证明了这种方法的有效性：LOM-4B 在本体补全任务中达到了 88.8% 的准确率，在复杂图推理任务中达到了 94% 的准确率，显著优于现有主流大型语言模型。这些结果指出了一种可扩展和确定性部署企业级人工智能的新范式。</p>
</div><p><strong>Moving Beyond the Parameter Race: Empowering AI to Build an Enterprise &#34;Business Logic Universe&#34;</strong></p>
<div class="lb-trans"><p><strong>超越参数竞赛：赋能人工智能构建企业 “商业逻辑宇宙”</strong></p>
</div><p>For years, the adoption of large models in enterprise scenarios has largely followed a linear assumption of &#34;more parameters lead to better performance&#34;. However, in real-world deployments, this approach often falls short, such as producing unstable reasoning, inconsistent outputs, and a persistent disconnect from business logic. The core reason is that traditional large models rely on probabilistic token prediction, without a structured understanding of enterprise systems, and even lacking the ability to autonomously construct a logic framework that fits the actual business needs of enterprises.</p>
<div class="lb-trans"><p>多年来，在企业场景中采用大型模型在很大程度上遵循了 “更多参数意味着更好性能” 的线性假设。然而，在实际部署中，这种方法往往不尽如人意，例如产生不稳定的推理、不一致的输出，以及与商业逻辑的持续脱节。核心原因在于传统大型模型依赖于概率性标记预测，缺乏对企业系统的结构化理解，甚至缺乏自主构建符合企业实际业务需求的逻辑框架的能力。</p>
</div><p>The core breakthrough of YonLOM lies in moving beyond mere parameter scaling and instead empowering AI with the ability of building up its own autonomous logic system. Much like a senior domain expert, YonLOM can organize business entities, attributes and their interrelationships from fragmented structured and unstructured data, forming a coherent enterprise ontology. The ontology functions as a &#34;business logic universe&#34; that aligns with the real operations within all reasoning tasks place of the enterprise via a structured, internally consistent environment, where all reasoning unfolds, fundamentally ensuring the determinacy of reasoning.</p>
<div class="lb-trans"><p>用友 LOM 的核心突破在于超越单纯的参数扩展，而是赋能人工智能构建自主的逻辑系统。就像一位资深领域专家，LOM 能够从零散的结构化和非结构化数据中组织商业实体、属性及其相互关系，形成一个连贯的企业本体。该本体作为一个与企业所有推理任务的真实操作相一致的 “商业逻辑宇宙”，通过一个结构化、内部一致的环境展开所有推理，从根本上确保推理的确定性。</p>
</div><p>In the ontology construction phase, LOM simultaneously processes structured data held in databases and unstructured textual documents. Through a multi-stage generation and validation pipeline, it converts scattered business information into a standardized, machine-interpretable ontology structure, while ensuring logical consistency through iterative verification. For example, from organizational hierarchies in human resources, to financial account linkages, to upstream-downstream dependencies across supply chains, YonLOM systematically reconstructs the underlying logic embedded in enterprise data for supporting the next-step operations explicit and actionable.</p>
<div class="lb-trans"><p>在本体构建阶段，LOM 同时处理存储在数据库中的结构化数据和非结构化文本文件。通过多阶段的生成和验证流程，它将分散的商业信息转换为标准化的、机器可解释的本体结构，同时通过迭代验证确保逻辑一致性。例如，从人力资源中的组织层级，到财务账户的关联，再到供应链中的上下游依赖关系，LOM 系统地重构了嵌入企业数据中的基础逻辑，以支持下一步操作的明确和可执行性。</p>
</div><p><strong>A Tripartite Integrated Architecture for In-depth Fusion of Semantics and Structure</strong></p>
<div class="lb-trans"><p><strong>语义与结构深度融合的三方集成架构</strong></p>
</div><p>While autonomous ontology construction forms the foundation of LOM, it is the precise alignment of semantics and structure that makes this &#34;logic universe&#34; truly reflect the realities of enterprise business. Enterprise knowledge features in both semantical diversity and structural complexity. Traditional models often process text understanding and structural reasoning separately, leading to information loss and reasoning inaccuracies.</p>
<div class="lb-trans"><p>虽然自主本体构建构成了 LOM 的基础，但语义与结构的精确对齐使得这个 “逻辑宇宙” 真正反映了企业业务的现实。企业知识在语义多样性和结构复杂性方面具有特征。传统模型通常将文本理解和结构推理分开处理，导致信息丢失和推理不准确。</p>
</div><p>Through the tripartite integrated CAR architecture of Construct-Align-Reason, LOM dynamically integrates these three procedures into a cohesive cognitive framework. During the alignment phase, the model leverages a graph-aware encoder and reinforcement learning to accurately match the semantically generated information with the constructed ontology structure. Which enables real-time mapping between abstract graph nodes and real business entities of enterprises, as well recognizing dynamic updates of the ontology at the same time. As new business knowledge and insights merges during the interactions with users, LOM adjusts the ontology structure accordingly, ensuring that this &#34;logic universe&#34; always evolves in line with the development of enterprise business.</p>
<div class="lb-trans"><p>通过构建 - 对齐 - 推理的三方集成 CAR 架构，LOM 动态地将这三个过程整合为一个统一的认知框架。在对齐阶段，模型利用图感知编码器和强化学习，准确地将语义生成的信息与构建的本体结构匹配。这使得抽象图节点与企业的真实商业实体之间实现实时映射，同时识别本体的动态更新。当与用户的交互中出现新的商业知识和见解时，LOM 相应地调整本体结构，确保这个 “逻辑宇宙” 始终与企业业务的发展保持一致。</p>
</div><p>In the reasoning phase, LOM discards probabilistic guesswork and instead uses the autonomously constructed ontology as immutable business rules to execute strict deterministic reasoning within the system. Whether performing complex graph algorithms such as shortest path and minimum spanning tree, or navigating multi-hop business relationships, LOM produces verifiable and reliable results. This logical reasoning execution transforms AI from simulated reasoning into practical business operations, perfectly adapting to enterprise scenarios where require extremely high accuracy such as finance, supply chain and production.</p>
<div class="lb-trans"><p>在推理阶段，LOM 摒弃了概率性猜测，而是使用自主构建的本体作为不可变的商业规则，在系统内执行严格的确定性推理。无论是执行最短路径和最小生成树等复杂图算法，还是导航多跳商业关系，LOM 都能产生可验证和可靠的结果。这种逻辑推理执行将人工智能从模拟推理转变为实际商业操作，完美适应需要极高准确性的企业场景，如金融、供应链和生产。</p>
</div><p><strong>Surpassing Mainstream Models in Real-World Tests: High Cognitive Density with Fewer Parameter</strong></p>
<div class="lb-trans"><p><strong>在真实世界测试中超越主流模型：更少参数实现高认知密度</strong></p>
</div><p>Genuine technological value brought by paradigm innovation is ultimately validated in real-world scenarios. To this end, Yonyou AI Lab conducted comprehensive testing on LOM using production data from diverse enterprise domains, including human resources, finance, assets management, manufacturing, and supply chain. The benchmark dataset included 19 graph reasoning tasks, comparing LOM's performance against that of mainstream large language models.</p>
<div class="lb-trans"><p>范式创新带来的真正技术价值最终在真实场景中得到了验证。为此，用友 AI 实验室对 LOM 进行了全面测试，使用来自人力资源、金融、资产管理、制造和供应链等多个企业领域的生产数据。基准数据集包括 19 个图推理任务，将 LOM 的性能与主流大型语言模型进行了比较。</p>
</div><p>Test results show that LOM-4B, with only 4 billion parameters, achieved an accuracy of 93% across all tasks on average. LOM-32B, with 32 billion parameters, further increased the accuracy to 94%, especially excelling in tasks requiring deterministic reasoning such as shortest path, cycle detection and minimum spanning tree. In contrast, mainstream models, though endowed with vastly larger parameter counts, demonstrated proficiency primarily in shallow semantic tasks, where surface-level patterns and associations are sufficient. However, they struggled when confronted with complex structural reasoning challenges, where deterministic logic and sophisticated relationships between business entities are deeply embedded as an intertwined graph. In these tasks, the model's performance deteriorated sharply, with some tasks registering near-zero accuracy, highlighting the inherent limitations of relying solely on probabilistic models for tasks that demand precise, rule-based reasoning and structural coherence.</p>
<div class="lb-trans"><p>测试结果显示，LOM-4B 仅有 40 亿个参数，在所有任务上的平均准确率达到了 93%。而 LOM-32B 则拥有 320 亿个参数，进一步将准确率提高至 94%，特别是在需要确定性推理的任务中表现出色，如最短路径、循环检测和最小生成树。相比之下，主流模型虽然参数数量大幅增加，但主要在浅层语义任务中表现出色，这些任务只需表面模式和关联即可。然而，当面临复杂的结构推理挑战时，它们却显得力不从心，因为这些挑战中确定性逻辑和商业实体之间复杂关系深深交织在一起。在这些任务中，模型的表现急剧下降，部分任务的准确率接近零，突显了仅依赖概率模型在需要精确、基于规则的推理和结构一致性的任务中的固有限制。</p>
</div><p>This result underscores a core insight: the true strength of enterprise-grade AI lies not in accumulating parameters, but in cognitive density. By merging the technical idea of neuro-symbolic fusion, LOM delivers enhanced logical reasoning capabilities with fewer parameters, demonstrating that for industrial AI, logical density is more valuable than parameter scale. For enterprises, this means accessing AI reasoning capabilities that are more aligned with business needs in a more stable and reliable way without investing huge costs in deploying ultra-large parameter models.</p>
<div class="lb-trans"><p>这一结果强调了一个核心见解：企业级人工智能的真正力量不在于参数的积累，而在于认知密度。通过融合神经符号融合的技术理念，LOM 以更少的参数提供了增强的逻辑推理能力，证明了对于工业人工智能而言，逻辑密度比参数规模更具价值。对于企业而言，这意味着可以以更稳定可靠的方式获取与业务需求更为契合的人工智能推理能力，而无需在部署超大参数模型上投入巨额成本。</p>
</div><p><strong>7D Logical Autonomy: Unleashing the Next Generation of Enterprise AI</strong></p>
<div class="lb-trans"><p><strong>7D 逻辑自主性：释放下一代企业人工智能的潜力</strong></p>
</div><p>Building on this breakthrough, Yonyou AI Lab has also developed a 10-Dimensional Cognitive Framework that outlines the evolution of AI models. According to this framework, current mainstream large models and agents are still in the 6D stage, capable of optimizing task execution paths but lacking the ability to autonomously construct underlying logic systems. This limitation is the primary reason for the reasoning challenges in enterprise applications.</p>
<div class="lb-trans"><p>在这一突破的基础上，用友 AI 实验室还开发了一个 10 维认知框架，概述了人工智能模型的演变。根据该框架，目前主流的大型模型和代理仍处于 6D 阶段，能够优化任务执行路径，但缺乏自主构建基础逻辑系统的能力。这一限制是企业应用中推理挑战的主要原因。</p>
</div><p>The 7D logical autonomy achieved by LOM marks a significant step for Enterprise AI to evolve to higher dimensions. It enables AI to construct the logical frameworks for reasoning from scratch, essentially empowering it with the ability to set the rules of the game, rather than just playing it. This evolution transforms AI from a data processing tool into a true knowledge expert and decision-making assistant of enterprise business logic, paving the way for future AI paradigms that are capable to make autonomous decisions and actively trigger business executions.</p>
<div class="lb-trans"><p>LOM 所实现的 7D 逻辑自主性标志着企业人工智能向更高维度演进的重要一步。它使人工智能能够从零开始构建推理的逻辑框架，实质上赋予其设定游戏规则的能力，而不仅仅是参与游戏。这一演变将人工智能从数据处理工具转变为真正的知识专家和企业业务逻辑的决策助手，为未来能够自主决策并主动触发业务执行的人工智能范式铺平了道路。</p>
</div><p>Today, as enterprise digital transformation has entered a more advanced phase, the full realization of data assets' value increasingly depends on the abilities in generating meaningful ontology base, adopting accurate sub-graph retrieval tools, and thus providing in-depth reliable insights for actions. YonLOM not only offers deterministic reasoning capabilities but also introduces a novel solution for integrating large models into industry practices.</p>
<div class="lb-trans"><p>如今，随着企业数字化转型进入更高级的阶段，数据资产价值的全面实现越来越依赖于生成有意义的本体基础、采用准确的子图检索工具，从而为行动提供深入可靠的洞察。YonLOM 不仅提供确定性推理能力，还为将大型模型整合到行业实践中引入了一种新颖的解决方案。</p>
</div><p>Looking ahead, the 7D logical autonomy of LOM lays the foundation for further advancements, Yonyou AI Lab will continue to evolve the model from a reasoning engine towards a dynamic, intelligent system capable for more reliable action execution advising and strategic decision-making.</p>
<div class="lb-trans"><p>展望未来，LOM 的 7D 逻辑自主性为进一步的进展奠定了基础，用友 AI 实验室将继续将模型从推理引擎演变为一个动态智能系统，能够提供更可靠的行动执行建议和战略决策。</p>
</div><p>For more detailed insights, you are welcomed to find the full preprint here: https://chinaxiv.org/abs/202603.00072</p>
<div class="lb-trans"><p>欲获取更详细的见解，欢迎访问完整的预印本： https://chinaxiv.org/abs/202603.00072</p>
</div><p>SOURCE Yonyou</p>
<div class="lb-trans"><p>来源：用友</p>
</div><p><img src="https://imageproxy.pbkrs.com/https://rt.prnewswire.com/rt.gif/query-RGF0ZUlkPTIwMjYwNDEwJk5ld3NJdGVtSWQ9Q04zMTQ1MyZUcmFuc21pc3Npb25fSWQ9MjAyNjA0MTAwMjQ4UFJfTkVXU19VU1BSX19fX19DTjMxNDUz" alt="" original-src="https://rt.prnewswire.com/rt.gif?NewsItemId=CN31453&amp;Transmission_Id=202604100248PR_NEWS_USPR_____CN31453&amp;DateId=20260410"/></p>

02:48 ETLOM：统一本体构建和语义对齐以实现大规模的确定性企业推理