文章背景与核心概要
在受策略规范的 AI 智能体应用中，如何在保证模型灵活解释特定案例证据的同时，强制其严格遵守预定的授权流程，一直是技术落地的核心挑战。本文介绍了一种名为 STAGE（Stateful Translation to Agentic Graph Execution with Policy-Scoped Context and Deterministic Control）的新型可执行图框架，旨在通过将程序化控制与大模型（LLM）的判断进行解耦，大幅提升受策略治理的智能体系统的可靠性。
STAGE 框架将大模型的推理限制在“策略限定节点”（policy-scoped nodes）内，而将整体的工作流逻辑委派给确定性的代码执行。在每个节点，模型接收与任务相关的策略上下文并返回类型化的结果，同时由协调器强制执行经过审查的执行契约。实验结果表明，与传统的单体全策略执行（monolithic full-policy execution）相比，STAGE 在不同程序复杂度的多步骤工作流（如银行纠纷和电信业务）中，均显著提高了任务成功率和多次运行的可靠性，特别是在深层 Telecom 和 Smart Dispute 工作流中取得了巨大的性能增益。
STAGE：基于策略限定上下文与确定性控制的有状态转换至智能体图执行 (STAGE: Stateful Translation to Agentic Graph Execution with Policy-Scoped Context and Deterministic Control)
arXiv: 2608.22538 [cs.AI]
Submitted: 23 Aug 2026
Authors: Mengxi Luo, Changjia Chen, An Cao, Zirong Huang, Wanyi Dai
摘要 (Summary)
STAGE 是一个可执行图框架，旨在增强受策略治理的 AI 智能体的可靠性。通过将程序化控制与模型判断解耦，STAGE 将大模型的推理限制在“策略限定节点”中，同时将工作流逻辑委派给确定性代码。该架构确保智能体在保持解释案例特定证据的灵活性同时，严格遵守授权程序。实证结果表明，任务成功率和可靠性得到了显式提升，尤其是在银行纠纷和电信等复杂的多步骤工作流中。
STAGE is an executable-graph framework designed to enhance the reliability of policy-governed AI agents. By decoupling procedural control from model judgment, STAGE confines the LLM's reasoning to "policy-scoped nodes" while delegating the workflow logic to deterministic code. This architecture ensures that agents adhere strictly to authorized procedures while maintaining the flexibility to interpret case-specific evidence. Empirical results demonstrate significant improvements in task success and reliability, particularly in complex, multi-step workflows like banking disputes and telecommunications.
摘要详述 (Abstract)
受策略治理的智能体必须在遵循授权程序的同时解释案例证据。我们提出了 STAGE，这是一个可执行图框架，它将模型的判断限制在策略限定节点中，同时将程序控制置于确定性代码中。
在每个节点，模型接收与任务相关的策略上下文并返回类型化的结果，而协调器则强制执行审查后的执行契约。我们在 SOP-Bench Referral Abuse、两个 \(\tau^2\)-bench 领域以及专属银行基准测试 Smart Dispute 上评估了 STAGE。与单体全策略执行相比，STAGE 通常提高了各种程序复杂度工作流中的任务成功率和重复运行的可靠性。最大的提升出现在更深的电信（Telecom）和智能纠纷（Smart Dispute）工作流中，根据模型的不同，\(\mathrm{Pass}^3\) 分别提高了 7.5--55.0 和 57.2--65.7 个百分点。这些结果表明，将策略限定上下文与确定性程序控制相结合，可以提高策略执行的可靠性。
Policy-governed agents must interpret case evidence while following an authorized procedure. We present STAGE, an executable-graph framework that confines model judgment to policy-scoped nodes while placing procedural control in deterministic code.
At each node, the model receives task-relevant policy context and returns a typed result, while the coordinator enforces the reviewed execution contract. We evaluate STAGE on SOP-Bench Referral Abuse, two \(\tau^2\)-bench domains, and Smart Dispute, a proprietary banking benchmark. Compared with monolithic full-policy execution, STAGE generally improves task success and repeated-run reliability across workflows of varying procedural complexity. The largest gains occur on the deeper Telecom and Smart Dispute workflows, where \(\mathrm{Pass}^3\) increases by 7.5--55.0 and 57.2--65.7 percentage points, respectively, depending on the model. These results show that combining policy-scoped context with deterministic procedural control can improve the reliability of policy execution.
访问论文 (Access Paper)
元数据 (Metadata)
| 字段 (Field) | 详情 (Details) |
|---|---|
| 学科 (Subjects) | 人工智能 (cs.AI) |
| DOI | 10.48550/arXiv.2608.22538 |
| 引用 (Cite as) | arXiv:2608.22538 [cs.AI] |
Field Details Subjects Artificial Intelligence (cs.AI) DOI 10.48550/arXiv.2608.22538 Cite as arXiv:2608.22538 [cs.AI]