Reproduce “KV-Cache Patching: Tracing How and Where Fine-Tuned Behaviors Arise” 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
Jonas Loos, Oliver Eberle
The authors test whether fine-tuned behavior transfers through cached prompt representations by swapping KV caches between LoRA fine-tunes and base models across three 4–8B models and 12 tasks. Response conventions and dispositions often move with the cache, while capabilities generally move less. In one Qwen2.5-7B medical-advice test, misalignment transfers at 4.8%, below the fine-tune’s 7.2%; findings are limited to these experiments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.