SwaT trains a small mapper so different language models can exchange KV caches instead of natural-language messages while keeping the underlying models frozen. The researchers evaluate 12 models from six families on 17 benchmarks. Experiments combining the method with cache compaction report roughly 40% average accuracy improvement and about fourfold gains on specific speed measures. That does not mean end-to-end application time universally falls by a factor of four. Deployment requires model-pair mappings and careful handling of cached state, with separate consideration of how to audit exchanges that are no longer human-readable messages.