aiqbench analysis
The paper argues that parallel token proposals in language-model decoding depend on the joint distribution across tokens, not just accurate per-token marginals. It builds joint laws from an integrable spin-chain model and shows that they can differ from their pairwise maximum-entropy approximations while matching one- and two-token moments. The results are finite-size mathematical existence proofs with no measured language-model or speedup gains, so readers should treat the practical relevance to speculative decoding as an open question.
A Bethe-Ansatz Source of Early Token Correlations: A marginal-preserving construction for parallel proposals
arXiv:2610.16781v1 Announce Type: new Abstract: Parallel token proposals in language-model decoding need a joint law, not merely accurate token marginals. We construct a family of coordinate-Bethe-ansatz laws from the periodic XXZ chain and compare each with its maximum-entropy pairwise projection at identical one- and two-site moments. In the eight-site, three-excitation sector, symmetry makes their entire probability-law difference one-dimensional, determined by a contiguous three-site moment. Along the Bethe branch, this gap varies with anisotropy, changes sign, and vanishes exactly at $\Delta=(3+\sqrt{3})/2$, where the eigenstate law becomes exactly pairwise. A nonlinear readout transfers a nonzero gap into the first two token positions while preserving their separate marginals. At $\Delta=1.5$ the deterministic binary token laws differ by 0.02654573 in total variation. Exact speculative verification then gives a strict accepted-prefix advantage for a constructed target and a neighborhood within targets sharing its first-token marginal. This is a mathematical existence result. It does not establish a language-model gain or wall-clock speedup.
Abstract:Parallel token proposals need a joint law, not merely accurate token marginals. We construct coordinate-Bethe-ansatz laws from the periodic XXZ chain and compare them with their maximum-entropy pairwise projections at identical one- and two-site moments. In an eight-site, three-excitation sector, symmetry makes the full probability-law difference one-dimensional, determined by a contiguous three-site moment $\delta$. Along the Bethe branch, this gap changes sign and vanishes exactly at $\Delta=(3+\sqrt3)/2$, where the eigenstate law becomes pairwise. A nonlinear two-token readout preserves separate token marginals yet exposes a distance $2|\delta|$; a direct three-token readout preserves all one- and two-token marginals yet exposes $4|\delta|$. At $\Delta=1.5$ these deterministic binary distances are $0.02654573$ and $0.05309145$. We give an exact target-alignment formula and prove a strict speculative accepted-prefix advantage throughout an unrestricted open neighborhood of a constructed target. These are finite-law existence results, not measured language-model or wall-clock gains.
6 pages, 2 tables. Ancillary files reproduce the numerical results
Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2610.16781 [cs.CL]
(or arXiv:2610.16781v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.16781 Focus to learn more arXiv-issued DOI via DataCite
high_precision_xxz.py
three_token_verification.json
verification.json
verify_alignment.py
verify_three_token.py
xxz_delta_sweep.json
xxz_delta_sweep.py
(3 additional files not shown) You must enabled JavaScript to view entire file list.
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 269 words
Related coverage

Expenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
METR · Jul 21, 2026

"Holy shit" graphs and how to get more of them
LessWrong · Oct 18, 2026

Large to Small Model Stitching Destroys What You Built It to Carry Along the Fit
LessWrong · Oct 11, 2026
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · Oct 8, 2026