1 paper · 1 filter
Hao Zhu, Phil Cuvin, Xinkai Yu +3
Agents are predominantly evaluated and optimized via task success metrics, which are coarse, rely on manual design from experts, and fail to reward intermediate emergent behaviors.…