Reproduce “To Cache or Not to Cache: Understanding and Mitigating Safety Incompatibility in Cross-Model KV Cache Reuse” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhiwei Lin, Jiahao Liu, Yan Liu, Chengpeng Wang
The authors evaluate how cross-model key-value cache reuse changes a receiver language model’s safety and introduce VReparo, which learns offline affine maps to align reused values with receiver-native values. Effects vary by model pair and reuse setting. Across four settings, VReparo lowers average HarmBench attack success rate from 57.28% to 39.21%, but scores remain worse than native behavior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.