跳到正文
arXiv · AI· Shuyang Xiang·· 10 小时前AI 评分35

A Conditional No-Hair Theorem for Transformers: How Language Models Forget Their Prompt

A Conditional No-Hair Theorem for Transformers: How Language Models Forget Their Prompt

AI 导读

研究指出,自回归 transformer 中 prompt 对 hidden states 的影响随生成 token 数 $t$ 呈幂律衰减 $t^{-\alpha}$。

来源:arXiv · AI · arxiv.org