KV-Cache Patching: Tracing How and Where Fine-Tuned Behaviors Arise

Patching the KV cache from a LoRA fine-tuned language model into its unmodified base model can transfer some fine-tuned behaviors, including responding in German and emergent misalignment, according to arXiv paper 2610.41480v1. The reverse patch, from base model to fine-tuned model, partially or fully removed fine-tuned behavior, and the effect grew with the share of the KV cache patched while concentrating at a few pivotal positions, with system-prompt and chat-template tokens having disproportionate influence. The findings indicate that fine-tuned behavior arises partly from altered processing of earlier tokens through attention, not only during generation of the answer tokens where the change appears.

arXiv:2610.41480v1 Announce Type: new Abstract: Fine-tuning a language model changes how it responds, but where is this fine-tuned behavior instantiated? To better understand this, we evaluate whether the change in model behavior arises only during the generation of the answer tokens where the change manifests, or whether altered processing of previous tokens causes it via the attention mechanism. For this, we patch the KV cache from a LoRA fine-tuned model into the unmodified base model, and vice versa. Our results show that for some behaviors, like responding in German or emergent misalignment, patching the KV cache can indeed transfer them from the fine-tuned to the base model. Correspondingly, patching the KV cache from base to fine-tune can partially or fully remove fine-tuned behavior. Furthermore, we find that the effect of patching grows with how much of the KV cache is patched and concentrates at a few pivotal positions, with system-prompt and chat-template tokens having disproportionate influence.