TGV-KV: Text-Grounded KV Eviction for VLMs

Arxiv CS.CV (computer vision)

The paper introduces TGV-KV, a text-grounded KV eviction method for VLMs that reduces memory consumption by selectively removing redundant tokens. It leverages cross-modal attention scores and achieves up to 50% cache reduction with minimal performance degradation. Evaluated on multiple benchmarks, TGV-KV demonstrates efficiency gains.

vision-language-models

Arxiv CS.CV (computer vision)·Jizhihui LiuRuizi HanMiao ZhangRui ShaoXuebo LiuWeili GuanYaowei Wang

8 days ago