Minghao Guo

Forget complex architectures: the secret to self-improving LLM agents lies in teaching them how to *interpret* their past failures, not just remember them.

LLMs still can't convincingly mimic human personas, especially when it comes to syntactic style and memory, despite advancements in other areas.