Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignment
The results show that task capability, additional reasoning, and consequence-oriented safety evaluation do not by themselves ensure authorization alignment, pointing to a tension between agent autonomy and preserving users'decision authority that calls for further human-centered study.
Zhi-Ping Zhang, Franklin Mingzhe Li, Zi-Hao Huang et al.
· 1 citation