Tag
Proposes treating semantic compression as a diffusion noise function for handling massive context beyond model windows, using multi-pass reading at decreasing compression levels. Untrained-model experiments show components work in isolation but the full chain needs training to resolve binding bottleneck.