文章背景与核心概要
本文由 MasterControl AI Lab 发布,探讨并评估了一种用于企业级数据分析的受控(Governed)方法。传统方法常允许语言模型在运行时同时处理规划与执行,而该论文提出的架构则实现了职责分离:由大语言模型负责理解用户意图,再由确定性策略(Deterministic Policy)选择并运行预先批准的分析程序。
通过对 440 次运行的严格评估,研究表明,采用策略治理的执行方式在维持严格的“答案与证据”契约方面,显著优于运行时规划智能体(Runtime-planning Agents)。该方案为企业数据分析提供了一个可靠、可重放且鲁棒的框架。
每次都能完美掌控十七 (MasterControl Seventeen Every Time)
摘要 (Summary)
本文由 MasterControl AI Lab 发布,评估了一种受控的企业分析方法。与允许语言模型在运行时同时处理规划与执行的做法不同,该架构将职责分离:大语言模型负责解释用户意图,而确定性策略则选择并运行预先批准的分析程序,该程序会同时返回结果和证据。
This paper, published by the MasterControl AI Lab, evaluates a governed approach to enterprise analytics. Rather than allowing a language model to handle both planning and execution at runtime, the proposed architecture separates concerns: an LLM interprets user intent, while a deterministic policy selects and runs pre-approved analytical programs.
我们证明,在关系运算、聚合、比较、窗口函数、排序和相似性等定义的分析类别内,这种限制依然能够保持足够的表达力。固定的含义、策略、数据以及执行规则,同时也使得结果具备可重放性。
Through an evaluation across 440 runs, the study demonstrates that policy-governed execution significantly outperforms runtime-planning agents in maintaining strict answer-and-evidence contracts, providing a reliable, replayable, and robust framework for enterprise data analysis.
论文元数据 (Paper Metadata)
- arXiv ID: arXiv:2609.03209 [cs.AI]
- 学科分类 (Subject): 人工智能 (
cs.AI) - 标题 (Title): 每次都能完美掌控十七 (MasterControl Seventeen Every Time)
- 作者 (Authors): MasterControl AI Lab
- 提交日期 (Submission Date): 2026年9月2日
- 主要许可 (Primary License): 知识共享署名 4.0
摘要正文 (Abstract)
我们研究了一种受控的企业分析方法:由语言模型负责理解问题,确定性策略选择并运行一个预先批准的分析程序,该程序会返回结果和证据。我们表明,在包含关系运算以及聚合、比较、窗口、排序和相似性的定义分析类别内,这种限制依然具有高度的表达力。固定的语义、策略、数据以及执行规则也使得分析结果能够被完全重放。
We study a governed approach to enterprise analytics: a language model interprets the question, while deterministic policy selects and runs a pre-approved analytical program that returns both results and evidence. We show that this restriction can remain expressive within a defined analytical class, using relational operations plus aggregation, comparison, windows, ranking, and similarity. Fixed meaning, policy, data, and execution rules also make results replayable.
在涵盖 440 次运行的测试中,三个 8B 参数模型在运行时生成 SQL 并选择工具,而 Qwen3-8B 仅用于解释意图,并由策略执行已批准的程序。在所有测试数据集上,330 次运行时规划(runtime-planning)实验无一完全匹配完整的“答案与证据”契约;而策略执行的分析器则实现了 110/110 的全中记录。这是一个特定配置下的结果,并非证明运行时智能体在其他设计下无法成功。
Across 440 runs, three 8B models generated SQL and selected tools at runtime, while Qwen3-8B interpreted intent only and policy executed the approved program. None of 330 runtime-planning episodes matched the full answer-and-evidence contract across all test datasets; the policy-executed analyzer matched 110 of 110. This is a configuration-specific result, not evidence that runtime agents cannot succeed under other designs.
访问与资源 (Access & Resources)
- 全文链接 (Full-Text Links): 查看 PDF | HTML(实验性) | TeX 源码
- 相关资产 (Associated Asset):
(查看许可证) - 外部引用 (External Citations): Google 学术 | Semantic Scholar | NASA ADS