Consistent agent behavior in CUGA

[evduester]

Consistent and reliable performance on tasks is a notorious problem for agents and it tends to get worse with growing task complexity. And CUGA agents are no exception to this general problem.

I am curious, are there any plans in CUGA to address consistent agent behavior ? For example by measuring and detecting inconsistencies. Once there is the ability to detect inconsistencies in agent trajectories one could image to leverage learning mechanisms to "learn to be consistent".

[aviyaeli]

Yes, consistency is a crucial ingredient for trustworthy agents. We are planning to invest in this area and welcome innovative ideas