今日速览
| 序号 | 标题 | 来源 | 日期 | 主题 | 推荐等级 |
|---|---|---|---|---|---|
| 1 | ToolPrivacyBench: Benchmarking Purpose-Bound Privacy in Tool-Using LLM Agents | arXiv | 2026-06-26 | AI-Agent | 高 |
| 2 | From Detection to Action: Using LLM Agents for Fault-Tolerant Control | arXiv | 2026-06-26 | RAG | 高 |
重点论文与技术动态
1. ToolPrivacyBench: Benchmarking Purpose-Bound Privacy in Tool-Using LLM Agents
- 来源:arXiv
- 日期:2026-06-26
- 作者/机构:Shijing Hu, Liang Liu, Zhu Meng, Zhicheng Zhao
- 主题标签:
AI-Agent,arXiv - 推荐等级:高
- 分类:cs.CR, cs.AI
一句话结论
ToolPrivacyBench正式定义了最小必要披露边界。
一段话:该基准通过轨迹级审计评估工具使用中的隐私过披露,基于2,150个案例的实验发现,任务成功执行并不等同于隐私披露适当,代理可能在中间工具调用中传输不必要的私有信息。
核心内容
- 传统函数调用基准仅评估任务完成和API正确性,隐私基准聚焦最终响应,均未捕捉多工具轨迹中的目的绑定信息流。
- ToolPrivacyBench审计任务私有原子是否仅路由至授权工具和下游接收点,同时评估任务完成与隐私过披露。
- 实验结果表明,成功工具执行常伴随隐私过披露,代理可能在中间调用中泄露非必要私有信息。
方法与数据
- 基准含2,150案例:1,150个完全合成的隐私敏感业务工作流和1,000个改编自现有基准的案例。
- 每个案例由策略知识库表示,通过比较工具参数与后端审计日志验证隐私合规性。
价值判断
- 值得关注:工具使用中隐私过披露的系统性评估需求。
- 可复用点:轨迹级审计方法和策略知识库框架。
- 局限/待核查:摘要未明确。
摘要
Large language models (LLMs) have increasingly moved from standalone text generation systems to agents that invoke external tools, access environments, and execute multi-step tasks. However, conventional function-calling benchmarks mainly evaluate task completion and API correctness, while privacy evaluation benchmarks typically focus on final responses or privacy judgments. Neither perspective captures purpose-bound information flow across an executed multi-tool trajectory. Motivated by this limitation in current agent evaluation, ToolPrivacyBench audits whether task-private atoms are routed only to authorized tools and downstream sinks, thereby evaluating both task completion and privacy over-disclosure during tool use. The benchmark contains 2,150 cases, including 1,150 fully synthetic privacy-sensitive business workflows and 1,000 cases adapted from existing multi-tool and function-calling benchmarks. Each case is represented by a policy knowledge base. After an agent executes against mock business backends, the evaluator compares recorded tool arguments and backend audit logs with this policy knowledge base. The evaluation covers nine widely used agents to characterize purpose-bound privacy over-disclosure. The results show that successful tool execution does not imply appropriate privacy disclosure: an agent may complete a task while transmitting unnecessary private information through intermediate tool calls. ToolPrivacyBench therefore formalizes a need-to-know disclosure boundary, under which each tool should receive only the information necessary for its stated purpose, and uses trajectory-level auditing to identify privacy over-disclosure in multi-tool workflows.2. From Detection to Action: Using LLM Agents for Fault-Tolerant Control
- 来源:arXiv
- 日期:2026-06-26
- 作者/机构:Javal Vyas, Milapji Singh Gill, Artan Markaj, Felix Gehlhoff, Mehmet Mercangöz
- 主题标签:
RAG,arXiv - 推荐等级:高
- 分类:eess.SY, cs.LG
一句话结论
该框架通过LLM代理实现故障检测到约束感知恢复动作的自动化,证明在离散和连续过程中使用轻量级模型能高效生成有效决策。
核心内容
- 构建多代理工作流分解操作员职责为监控、规划、动作合成、模拟、验证和重提示。
- 集成数字化工厂孪生体(DPPT)和Graph RAG层,基于CPSMod本体组织植物知识支持关系感知多跳检索。
- 在离散批混合模块和CSTR模拟中,轻量级LLM(GPT-4o-mini)在兼容过程动态的延迟预算内推导有效恢复决策。
方法与数据
- 方法:多代理工作流、数字化工厂孪生体(DPPT)、Graph RAG层。
- 数据:模拟评估基于离散批混合模块和连续搅拌釜反应器(CSTR)基准测试。
价值判断
- 值得关注:实现故障检测到恢复动作的端到端自动化,显著提升工业控制系统的故障容错能力。
- 可复用点:Graph RAG知识组织架构和多代理工作流可迁移至其他工业控制系统场景。
- 局限/待核查:摘要未明确提及局限。