In autoregressive language-model decoding, generating several token proposals in parallel requires knowing how the proposals relate to one another. A new arXiv preprint (arXiv:2610.16781) argues that matching individual token probabilities is not enough; the proposals need a joint law.
To supply such a law, the authors construct a family of coordinate-Bethe-ansatz distributions derived from the periodic XXZ spin chain. The title points to early token correlations, and the abstract frames the construction as marginal-preserving.
The paper then compares each constructed law with its maximum-entropy counterpart. Because the abstract is truncated, the announcement does not disclose the outcome of that comparison or the empirical implications.