works on
long-context memory 1model efficiency 1retrieval-augmented generation 1self-distillation 1transformer depth division 1

From the 1 of 3 linked papers with an AI index.

collaborators
Showing cs.CLShow all

1 paper · 1 filter

cs.CL2026

Understanding Is Done Early: A Depth Division of Labor in Large Language Models and Its Use for Unbounded-Context Memory

Hanzuo Liu, Xuan Qi, Chunyu Liu +6

The paper proposes CoMem, a method that stores intermediate transformer layer states as memory to enable efficient long‑context retrieval, showing that using lower‑mid layers for c…