paper web signal

LatentPress: 4.2M-Param Adapter Compresses Context 7.7x and Beats Raw-Context QA Quality

Summary

Passes a meaningful quality threshold: at 7.7x compression, LatentPress's continuous memory tokens produce higher downstream QA accuracy than reading the uncompressed original context—achieved with a frozen base model and an adapter under 0.1% its size, writing in 43ms.