ServiceNow introduces MosaicLeaks, a benchmark evaluating privacy leakage in deep research agents that combine private documents with external tools. The study reveals that current models frequently leak sensitive data, and optimizing solely for task performance exacerbates this risk. Their proposed PA-DR reinforcement learning method improves strict chain success to 58.7% while significantly reducing information leakage.
Termsslug: multi-hop-reasoning
Multi-hop Reasoning
多跳推理
A reasoning process requiring the connection of multiple dispersed facts or steps to derive a final answer.
Curated Intelligence
Hugging Face BlogT1·Agents
76 pts
Google ResearchT1·Agents
58 pts
Google has integrated 'Agentic RAG' into the Gemini Enterprise Agent Platform to address the limitations of traditional single-step RAG in handling multi-source, multi-hop queries. By enabling AI to autonomously plan, reason, and iteratively interact with disparate data silos, this approach significantly improves answer reliability and accuracy for complex business workflows.