Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse
The paper claims LLMs in the same family can hand off KV caches with a fitted linear map, avoiding much of the prefill cost.
The method targets matched KV shapes, selecting predictive source layers, removing RoPE from keys, and fitting per-head ridge regressions on a small calibration set. On Qwen3 14B to 32B, multiple source layers explained 79% of target key variance and 65% of value variance. Across six model pairs, four retained 73-98% of standalone-prefill accuracy, while two failed badly until a nonlinear MLP improved retention. The reported transfer ran 2.7-25x faster than recomputing prefill and stayed stable across multi-turn handoff. ArXiv · AI/CL/LG's note
The method targets matched KV shapes, selecting predictive source layers, removing RoPE from keys, and fitting per-head ridge regressions on a small calibration set. On Qwen3 14B to 32B, multiple source layers explained 79% of target key variance and 65% of value variance. Across six model pairs, four retained 73-98% of standalone-prefill accuracy, while two failed badly until a nonlinear MLP improved retention. The reported transfer ran 2.7-25x faster than recomputing prefill and stayed stable across multi-turn handoff. ArXiv · AI/CL/LG's note
score 6