Demonstrates the vulnerability of tool selection in LLM-derived agents, highlighting fairness issues in distribution and competition for external tools.
Key Points
Tool selection bias increased from 20% to 81%, showing vulnerability to manipulation of tool names and descriptions.
Defenses like paraphrasing were evaluated to mitigate bias in selection processes among various agents.
Emerging tool ecosystems face significant risks to fairness and competition due to biases in tool selection methods.
Results underscore the importance of security measures in the operation of LLM-based agents with external tools.