Long context compression

1 paper across 1 session

Poster Session 4

Thursday, December 4, 2025 · 4:30 PM → 7:30 PM

UniGist: Towards General and Hardware-aligned Sequence-level Long Context Compression

#3516 · Chenlong Deng, Zhisong Zhang, Kelong Mao, Shuaiyi Li, Tianqing Fang, Hongming Zhang, Haitao Mi, Dong Yu, Zhicheng Dou

We introduce UniGist, a unified gist token-based long context compression method without chunk-wise training, which significantly enhances long context retention and efficiency through a hardward-aligned design.