Topic
llm agents
New Research Reveals LLM Agents Often Choose Over-Privileged Tools, Posing Security Risks
A new study introduces ToolPrivBench to evaluate over-privileged tool selection in LLM agents. The research finds that agents commonly choose higher-privilege tools even when lower-privilege alternatives are sufficient, and that safety alignment does not prevent this. A privilege-aware post-training defense is proposed to reduce unnecessary high-privilege tool use.
Uncertainty Decomposition Enables LLM Agents to Proactively Seek Clarification on Ambiguous Tasks
A new arXiv paper introduces a prompt-based uncertainty decomposition method that separates action confidence from request uncertainty, enabling LLM agents to ask for clarification when task specifications are ambiguous. Evaluated on two new benchmarks with 50% underspecified tasks, the method improves clarification F1 by 73% over ReAct+UE and 36% over Uncertainty-Aware Memory across five LLM backbones.
New Survey Maps How Evidence Tracing and Execution Provenance Can Make LLM Agents Trustworthy
A new survey from arXiv explores evidence tracing and execution provenance as key mechanisms for ensuring trustworthiness in LLM-based agents. The paper defines a unified framework connecting retrieval grounding, tool-use safety, memory lineage, and failure diagnosis, and reviews benchmarks and open challenges.
New MBABench Evaluates LLM Agents on End-to-End Finance Spreadsheet Tasks
MBABench, a new benchmark from researchers, evaluates LLM agents on end-to-end spreadsheet tasks in finance, focusing on modeling and scenario analysis. The benchmark assesses accuracy, formula use, and formatting. Claude family models lead but still fall short of professional standards.