Give every word a place,
then let it spread.
LOTU puts tokens on a lattice. Each word is hashed to a cell and becomes a source of probability. The field then diffuses, decays and is topped up by the sources every step, and the picture shows where the mass collects. It is a way to see the lattice mechanics. It is not a model and it does not score real text.
One equation,
three moves.
The lattice holds a probability density P. Each step it changes by a diffusion term that spreads mass to neighbouring cells, a decay term that removes a fraction, and source terms where tokens sit:
∂P/∂t = D∇²P − λP + ΣSᵢ(z, t)
D is the diffusion coefficient (0.2 here), λ the decay rate (0.035) and Sᵢ the source for token i, scaled by its confidence.
Token anchoring
A word is hashed to a coordinate, and that cell becomes a source. Two words that hash close together reinforce each other, and a lone word leaves a single hot spot with a halo that widens as the steps run.
What the readouts mean
Mass is the sum of P over the lattice, so it rises while sources feed it and levels off when decay balances them. Mean source confidence is the average of the per-token values the hash and temperature produce. Higher temperature lowers every source, which lowers the plateau.