10 月 22 日，蚂蚁数科发布决策模型 LingDT-Decision-1.0，提供 4B 和 27B 两个参数规模版本，主要面向智能客服、安全风控、金融等企业应用场景。与侧重文本生成的通用大语言模型不同，该模型主要处理意图识别、任务分类、工具选择及风险判断等结构化决策任务，目前已上线蚂蚁数科 DTMaaS 平台，开发者可限时免费体验。
根据蚂蚁数科公布的测试结果，在部分企业业务场景中，LingDT-Decision-1.0 相比所选通用大语言模型，Token 消耗降低超过 90%，任务处理速度最高提升约 10 倍。在覆盖 23 项开源基准的评测中，其 27B 版本的综合评分超过了 Microsoft-Decision-1、Jev 1.13.0 及 Fastino/GLiDE 等决策模型。
减少文本生成环节，面向高频决策任务优化
在企业业务系统中，并非所有 AI 任务都需要生成完整的自然语言回答。大量实际需求是对输入信息进行识别、分类，再根据判断结果执行后续操作。
例如，在智能客服场景中，当用户提出“订单三天还没发货，请尽快处理”时，系统首先需要判断用户意图是催单、退款还是投诉，随后确定工单优先级、处理部门以及是否需要人工介入。这类任务通常具有明确的候选选项，更看重判断准确率、处理时延和调用成本。
通用大语言模型同样能够处理这些任务，但在一些实现方式中，需要先生成文本或结构化答案，再由业务系统解析和校验。对于调用频率较高、选项相对固定的业务，这一过程可能增加 Token 消耗和处理时间。
LingDT-Decision-1.0 采用了不同的处理思路，将部分需要生成答案再解析的任务，转变为在预设候选选项中直接进行判断，从而减少不必要的文本生成及后处理环节。
蚂蚁数科披露，目前该模型已在智能客服、安全风控及金融应用等业务中开展测试。其中，在数据标注场景中，相比 DS-V4-flash，LingDT-Decision-1.0 的 Token 消耗降低约 97%，任务处理速度提升约 10 倍。在智能客服场景中，相比 Qwen-3.7-flash，Token 消耗降低约 90%，处理速度提升约 1 倍，风险识别准确率由 76%提高至 89%。
金融应用方面，测试覆盖意图识别、工具调用和幻觉检测等任务。
蚂蚁数科数据显示，模型在相关任务中的处理速度提升约 5 倍，工具组合完全匹配率提高约 2.6 倍，幻觉召回率由 44%提升至 70%。
支持多选和按序决策，处理复合业务需求
除常见的单选、分类和评级任务外，LingDT-Decision-1.0 还增加了原生多选和按序决策能力，用于处理包含多个目标或存在执行依赖关系的业务请求。
以智能客服为例，假设用户提出：“我要退款、关闭续费、补开发票，但不要注销账号。”
对于只支持单一意图分类的系统而言，如果将整段请求归入退款类别，就可能遗漏关闭续费和补开发票等其他需求。
LingDT-Decision-1.0 的多选能力允许同一请求同时匹配多个候选结果，使系统能够识别复合意图，并为后续工具调用提供判断依据。
按序决策则进一步处理任务之间的依赖关系。针对上述请求，系统不仅需要识别多个操作，还需要判断哪些任务可以独立执行，哪些任务必须等待前置条件满足。例如，退款前需要完成订单核验，发票处理可能需要参考退款结果，而关闭自动续费可能不依赖退款流程。
这一功能采用有向无环图（DAG）的任务依赖关系表示方式，主要面向多步骤决策场景。
不过，识别任务依赖关系并不等于模型能够独立完成全部业务操作。对于涉及资金、账户权限及风险控制的任务，仍需要与企业现有业务规则、权限系统及执行工具配合。
23 项开源基准测试，27B 版本综合评分超过多款决策模型
在模型能力评测方面，蚂蚁数科公布了 LingDT-Decision-1.0 与其他决策模型的对比结果。
据其介绍，相关评测覆盖 23 项开源基准，采用单选评分占 90%、多选评分占 10%的加权方式计算综合成绩。在这一评测设置下，LingDT-Decision-1.0 的 27B 版本综合评分位居参评模型首位，高于 Microsoft-Decision-1、Jev 1.13.0 及 Fastino/GLiDE 等模型。
具体而言，在多选任务测试中，LingDT-Decision-1.0 的 4B 和 27B 版本准确率均超过包括 Microsoft-Decision-1、Jev 1.13.0 及 Star-lux 在内的 16 款参评决策模型。单选任务方面，该模型也取得了相对靠前的成绩。
近期，面向固定选项判断的决策模型开始受到关注。微软 Microsoft-Decision-1、Jev 等模型均采用了与通用文本生成不同的技术路线，尝试通过直接判断候选结果来降低部分决策任务的时延和调用成本。
这类模型与通用大语言模型并不完全构成替代关系。前者侧重解决候选范围明确、结果能够结构化表达的任务，后者则适用于开放式问答、复杂内容生成和需要灵活推理的场景。
对于企业而言，决策模型的潜在价值在于将部分高频、重复性的判断任务从通用大语言模型中分离出来，减少生成式处理的开销。但在真实生产环境中，除了分类准确率，还需要考察模型对模糊请求、异常输入、错误选项及业务规则变化的适应能力。
但无论如何，LingDT-Decision-1.0 的发布，还是为企业高频决策任务提供了新的模型选择。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
On 22 October, LingDT-Decisión-1.0, a decision model of publish, an ant count section, provided a version of the 4B and 27B parameters, mainly for applications in smart passenger service, safety management, finance, etc. Unlike the text-based generic large-linguistic model, which deals mainly with structured decision-making tasks such as intention identification, job classification, tool selection and risk judgement, it is now available on-line at the DTMaaas platform, where developers can experience limited-time free of charge.
According to the test results published by the Ant Numbers Section, LingDT-Decision-1.0 in some of the enterprise business scenarios reduced the consumption of Token by more than 90 per cent and the speed of task processing by about 10 times. In the evaluation covering 23 open-source baselines, the 27B version has a combined rating that exceeds decision models such as Microsoft-Decision-1, Jev 1.13.0 and Fastino/GLIDE.
Reduction of text generation links and optimization of HF decision-making tasks
In the business system, not all AI missions need to generate complete natural language answers. A large amount of actual demand is for input information to be identified, classified and followed up on the findings.
For example, in the context of smart passenger service, when a user proposes “orders for three days not to be delivered, please process them as soon as possible”, the system first needs to determine whether the user intends to press, refund or complain, and then determine the priority of the work order, the processing department and whether manual intervention is required. Such missions usually have clear candidate options, with greater emphasis on judging accuracy, processing latency and moving costs.
The generic large-language model can also handle these tasks, but in some ways it will require textual or structured answers to be analysed and validated by the operational system. This process may increase token consumption and processing time for more frequent operations with relatively fixed options.
LingDT-Decisión-1.0 takes different approaches, shifting part of the task of generating answers for rectification to a direct judgement in the preset candidate option, thereby reducing unnecessary text generation and reprocessing.
The Ant Numbers Section has disclosed that the model is currently being tested in operations such as smart passenger service, security wind control and financial applications. Of these, token consumption was reduced by about 97 per cent and task processing speed by about 10 times compared to DS-V4-flash, LingDT-Decision-1.0. In the smart passenger service scenario, token consumption was reduced by about 90 per cent compared to Qwen-3.7-flash, processing speed increased by about 1 per cent and risk recognition accuracy increased from 76 per cent to 89 per cent.
For financial applications, the tests cover tasks such as intention identification, tool call and hallucinogenic detection.
Data from the ants section show that the processing speed of the model in the relevant task is about five times higher, the full match of the tool combination is about 2.6 times higher, and the hallucinogenic recall rate is up from 44 per cent to 70 per cent.
Support for multiple selection and sequenced decision-making to address complex business needs
In addition to common individual, classification and rating assignments, LingDT-Decisión-1.0 also increases the original multiple selection and sequential decision-making capacity to handle business requests that contain multiple objectives or have an enforcement dependency relationship.
In the case of smart passenger service, it is assumed that the user proposes: “I want to refund the money, close the money, supplement the bill, but do not cancel the account”.
For systems that support only a single intended classification, if the entire paragraph request is placed in the category of refunds, other requirements, such as closing fees and billing, may be omitted.
The multi-selection capability of LingDT-Decision-1.0 allows multiple candidate outcomes to be matched with the same request, enables the system to identify composite intent and provides a basis for judgement for subsequent tool call.
Sequenced decision-making further addresses the dependency between mandates. In response to the above requests, the system needs not only to identify multiple operations, but also to determine which tasks can be carried out independently and which must wait until the conditions are met. For example, the order verification is completed before refunds are made, invoice processing may require reference to the refund results, and closure of automatic renewal fees may not depend on the refund process.
This function is expressed in terms of relationship to the tasks of the uncharted map (DAG) and is primarily oriented towards a multi-step decision-making landscape.
However, the recognition of mission dependency does not mean that the model is capable of performing the entire business independently. Tasks involving funds, accounts permission and risk control still need to be aligned with the existing business rules of the enterprise, the permission system and implementation tools.
23 open source baseline tests, 27B version composite rating exceeding multiple decision model
In modelling capacity assessment, the ants count section published a comparison of LingDT-Decision-1.0 with other decision models.
According to it, the assessment covered 23 open-source baselines, with a weighting of 90 per cent for single scores and 10 per cent for multiple scores. Under this evaluation setting, the 27B version of LingDT-Decisión-1.0 was ranked first in the reference model, higher than the Microsoft-Decisión-1, Jev 1.13.0 and Fastino/GLIDE models.
Specifically, during the multi-selection test, the 4B and 27B versions of LingDT-Decision-1.0 exceeded the 16-year participatory decision-making model, including Microsoft-Decision-1, Jev 1.13.0 and Star-lux. In the case of individual assignments, the model has also achieved relatively advanced results.
In the near future, attention has begun to be paid to decision-making models that are oriented towards fixed options. Models such as Microsoft Microsoft-Decisión-1 and Jev have used different technical routes from generic text generation to attempt to reduce latency and call costs for some decision-making tasks by directly judging candidate results.
Such models do not fully constitute an alternative relationship to the generic large language model. The former focuses on addressing mandates with clear ranges of candidates and results that can be structured, while the latter applies to open-ended questions and answers, complex content generation and scenarios requiring flexible reasoning.
For the enterprise, the potential value of the decision-making model is to separate some of the high frequency, repetitive judgement tasks from the generic large language model to reduce the cost of generating processing. In the real production environment, however, in addition to the classification accuracy rate, the adaptability of models to vague requests, unusual inputs, error options and changes in business rules needs to be examined.
In any event, LingDT-Decisión-1.0 publish provided a new model selection for high-frequency enterprise decision-making tasks.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.