OVAL: Output-Aware Local Page Bases for KV Cache Retrieval
The paper’s claim is that KV page retrieval should optimize for the attention output, not just page relevance.
OVAL builds page encodings from the joint structure of keys and values, while keeping the key information needed for retrieval. It is training-free and does not add value-dependent inference-time statistics. The authors say it stores the same-size representation and has the same decode-time scoring cost as a key-only spectral baseline. Across long reasoning, long-context understanding, and long-generation benchmarks, it improves on that baseline and is competitive with recent KV cache compression and retrieval methods. ArXiv · AI/CL/LG's note
OVAL builds page encodings from the joint structure of keys and values, while keeping the key information needed for retrieval. It is training-free and does not add value-dependent inference-time statistics. The authors say it stores the same-size representation and has the same decode-time scoring cost as a key-only spectral baseline. Across long reasoning, long-context understanding, and long-generation benchmarks, it improves on that baseline and is competitive with recent KV cache compression and retrieval methods. ArXiv · AI/CL/LG's note
score 5