Where Should a Document Live: Context, Representations, or Parameters?
KV-cache “Cartridges” beat fine-tuning-style document injection in the paper’s controlled tests.
The authors compare context-window documents, KV-cache representations, and parametric adaptation across five knowledge-heavy benchmarks. In oracle settings, Cartridges are reported as the most accurate method at nearly every storage budget, about 10 points ahead of parametric methods. In multi-document retrieval, Cartridges are the only method said to match in-context learning, but they also show catastrophic forgetting on control benchmarks. Source: ArXiv · AI/CL/LG's note.
The authors compare context-window documents, KV-cache representations, and parametric adaptation across five knowledge-heavy benchmarks. In oracle settings, Cartridges are reported as the most accurate method at nearly every storage budget, about 10 points ahead of parametric methods. In multi-document retrieval, Cartridges are the only method said to match in-context learning, but they also show catastrophic forgetting on control benchmarks. Source: ArXiv · AI/CL/LG's note.
score 5