For top-100-page sparse-context prediction, recall declined quickly with increasing token lag under temporal reuse, degraded more slowly with cross-layer offset, and improved from same-token refresh layer 0 to layer 2 before changing little. The fixed source—penultimate layer of the previous token—was reported to attain comparable recall while requiring only one full-attention layer per draft step. · CiteArk
For top-100-page sparse-context prediction, recall declined quickly with increasing token lag under temporal reuse, degraded more slowly with cross-layer offset, and improved from same-token refresh layer 0 to layer 2 before changing little. The fixed source—penultimate layer of the previous token—was reported to attain comparable recall while requiring only one full-attention layer per draft step.
