Do LLMs Benefit From Their Own Words?

Jenny Y. Huang, Leshem Choshen, Wei Sun +6

In multi-turn conversations, large language models typically condition on the full conversation history: both past user prompts and assistant responses. We revisit this design choi…