CONDUIT: A Unified Residual-Stream Restoration Framework for KV Cache Reuse in Vision-Language Models
CONDUIT, a training-free refresh policy that unifies single- and multi-image reuse as residual-stream restoration, is proposed, building on norm-weighted attention and an accessible pre-output cached-value-norm proxy, then applies empirical image-level relevance amplification before one global selection.
Pengan Chen, Kai-Sheng Zheng, Liang Hong et al.
· 0 citations