实验结果相当令人印象深刻。REFRAG在大多数情况下实现巨大加速且准确性无损。在超长上下文的16倍压缩（k=16）下，REFRAG的TTFT比LLaMA快约16.5倍。k=32时TTFT达到约32.9倍LLaMA（≈30.85倍报告值），与论文声称的30.85倍加速基本吻合。困惑度和下游准确性基本保持不变。

Hulkenberg has left Formula 1 twice - and twice he's come back. The first time, he was a promising rookie shuffled out of Williams at the end of 2010, returning a year later with Force India. The ...

Mary Brunkow, Fred Ramsdell and Shimon Sakaguchi are awarded the Nobel Prize in Physiology or Medicine 2025 for their ...