Abstract
Multi-codebook generative retrieval produces semantic identifiers by autoregressively generating K codebook tokens. The disclosure describes globally coordinated decoding by formulating the K-step process as a sequentially composed optimal transport problem with K transport plans linked by marginal consistency constraints. A retrieval transport cost is defined per step from decoder conditional probabilities and may include a learned future value term and a diversity penalty. An entropically regularized objective is solved using composed Sinkhorn iterations that perform forward and backward scaling passes across steps to satisfy source, intermediate, and target marginals. A global beam computation budget is adaptively allocated across steps based on mass distributions or entropies of computed output marginals, and low-mass candidates are pruned using transport-plan support. The resulting transport plans are decoded to output a set of multi-token semantic identifiers. Integration modes include full replacement of standard beam search, post-hoc refinement, and hybrid operation based on per-step uncertainty.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Composed OT Beam Search", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10638