Research questionHow can long-context language models reduce per-token reads from million-token KV caches without losing accuracy?With million-token conversations, each generated token may require scanning the full KV cache, making decoding expensive. Proxy-based selection can reduce what is read but still adds linear work at every step.