I believe that, absolutely. I just can't see any examples that exercise this outside of toy examples. (And I have severe doubt LLMs will reliably evaluate those

I believe that, absolutely. I just can't see any examples that exercise this outside of toy examples. (And I have severe doubt LLMs will reliably evaluate those complex examples by themselves. You're better off adding memory and agentic behavior via a formal language wrapper)