文章背景与核心概要
自动化数据科学通常涉及解决结构化的模型选择问题,包括选择数据转换、特征表示、网络架构、训练流程、评估协议以及优化策略。传统的 AutoML 系统虽然能自动化部分流程,但往往受限于预定义的模型和超参数搜索空间;而大语言模型(LLM)智能体虽然通过检索、代码生成和执行反馈提供了灵活性,但其建模决策往往缺乏结构、难以验证且不易复用。
为了克服这些局限,本文推出了 MOSAIC(Modular Orchestration for Structured Agentic Intelligence and Composition,面向结构化智能体与组合的模块化编排)框架。该框架引入了一种基于内存支撑的模型选择和工作流构建的结构化智能体方案,包含语义任务画像、蓝图构建、基于证据的代码生成以及验证与优化四大核心步骤。该研究在金融时间序列预测与生成任务中进行了实例化应用,实验结果证明其在任务性能、执行成功率以及决策可追溯性方面均有显著提升。
MOSAIC: Modular Orchestration for Structured Agentic Intelligence and Composition
arXiv: 2606.00708 [cs.AI]
Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Submitted: 30 May 2026 (v1); Last revised: 23 Aug 2026 (this version, v2)
DOI: 10.48550/arXiv.2606.00708
arXiv: 2606.00708 [cs.AI]
Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Submitted: 30 May 2026 (v1); Last revised: 23 Aug 2026 (this version, v2)
DOI: 10.48550/arXiv.2606.00708
Authors
Yifan Bao, Xinyu Xi, Xinyu Liu, Wen Ge, Lei Jiang, Kevin Zhang, Raad Khraishi, Yihao Ang, Anthony K.H. Tung, Lukasz Szpruch, Hao Ni
Authors
Yifan Bao, Xinyu Xi, Xinyu Liu, Wen Ge, Lei Jiang, Kevin Zhang, Raad Khraishi, Yihao Ang, Anthony K.H. Tung, Lukasz Szpruch, Hao Ni
Abstract Summary
Automated data science typically involves solving structured model-selection problems by choosing data transformations, feature representations, architectures, training procedures, evaluation protocols, and refinement strategies.
- Limitations of Traditional AutoML: Systems automate parts of this pipeline, but are usually constrained to predefined search spaces for models and hyperparameters.
- Limitations of LLM Agents: While offering flexibility via retrieval, code generation, and execution feedback, their modeling decisions often remain unstructured, difficult to verify, and hard to reuse.
- The MOSAIC Solution: MOSAIC (Modular Orchestration for Structured Agentic Intelligence and Composition) introduces a structured agentic framework for memory-grounded model selection and workflow construction.
- Semantic Task Profiling: Analyzes the task and dataset.
- Blueprint Construction: Retrieves prior cases and source-code modules to build an intermediate representation detailing components, composition, interface constraints, and execution requirements.
- Grounded Code Generation: Grounds LLM code generation in retrieved evidence rather than unconstrained synthesis.
- Validation & Refinement: Validates candidate models via execution and refines them using diagnostic feedback, training traces, task metrics, and a failure-aware reinforcement learning policy.
- Domain Application: Instantiated on financial time-series forecasting and generation, where models must balance predictive accuracy, distributional fidelity, execution reliability, and complex financial criteria (e.g., risk and tail behavior). Experimental results demonstrate improvements in task performance, execution success, and decision traceability.
Abstract Summary
自动化数据科学通常涉及通过选择数据转换、特征表示、网络架构、训练流程、评估协议以及优化策略来解决结构化的模型选择问题。
- 传统 AutoML 的局限性: 尽管系统能够自动化流水线的某些部分,但通常受限于预定义的模型和超参数搜索空间。
- LLM 智能体的局限性: 尽管它们通过检索、代码生成和执行 feedback 提供了灵活性,但其建模决策往往仍然缺乏结构、难以验证且不易复用。
- MOSAIC 解决方案: MOSAIC(Modular Orchestration for Structured Agentic Intelligence and Composition)引入了一种用于基于内存支撑的模型选择和工作流构建的结构化智能体框架。
- 语义任务画像(Semantic Task Profiling): 分析任务和数据集。
- 蓝图构建(Blueprint Construction): 检索先前的案例和源代码模块,以构建详细描述组件、组合、接口约束和执行要求的中间表示。
- 基于证据的代码生成(Grounded Code Generation): 将 LLM 代码生成建立在检索到的证据之上,而不是进行不受约束的合成。
- 验证与优化(Validation & Refinement): 通过执行验证候选模型,并利用诊断反馈、训练轨迹、任务指标以及具备防错意识的强化学习策略对其进行优化。
- 领域应用: 在金融时间序列预测与生成任务中进行了实例化,其中模型必须平衡预测准确性、分布保真度、执行可靠性以及复杂的金融标准(例如风险和尾部行为)。实验结果证明了其在任务性能、执行成功率以及决策可追溯性方面的提升。
Document Links & Resources
- Full-Text Access:
- View PDF
- HTML (Experimental)
- TeX Source
- License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (
view license)
Document Links & Resources
- 全文访问:
- 查看 PDF
- HTML(实验性)
- TeX 源码
- 许可证: 知识共享 署名-非商业性使用-禁止演绎 4.0 国际许可协议 (
查看许可证)
External References & Tools
- Citations: NASA ADS | Google Scholar | Semantic Scholar
- Associated Code & Demos: Hugging Face | CatalyzeX Code Finder | Replicate
External References & Tools
- 引用: NASA ADS | Google Scholar | Semantic Scholar
- 相关代码与演示: Hugging Face | CatalyzeX Code Finder | Replicate