Re-computing the whole conversation every time you switch models? That might be wasted work. Title: Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse URL: ❓ What does this paper actually do? 💡 NVIDIA researchers show how to "translate" the KV cache computed by one model size and reuse it directly on a different-sized model from the same family, using a training-free, closed-form ridge regression mapping. ❓ Why bother transferring across model sizes? 💡 Production systems escalate from small to large models or swap models mid-conversation for cost/quality reasons. Recomputing the whole context from scratch every time (re-prefill) wastes a lot of compute. ❓ Does accuracy hold up? 💡 On good pairs it retains 73-98% of the target model's standalone accuracy. Weak pairs degrade badly, but adding a small MLP recovers up to 36.8 points of HellaSwag accuracy. ❓ How much faster is it? 💡 2.7-25x faster than re-prefill, with the biggest gains (25x) at 32K-token contexts, and accuracy stays nearly stable across 10-turn conversations. #LLM# #KVCache#