Abstract
Techniques are described for multi-objective beam search in generative retrieval systems that autoregressively generate semantic identifiers across multiple codebook steps. Each beam carries a vector of objective scores, such as relevance, diversity, freshness, and business value. At each decoding step, candidate beam extensions are evaluated and pruned by Pareto dominance using non-dominated sorting to form fronts. Beam selection proceeds by taking fronts in order until a beam budget is reached, and crowding distance is used to select among candidates within a front to promote spread across objective space. In some implementations, each beam also carries a preference vector used to bias expansion scoring while retaining the full objective score vector for Pareto pruning, and the set of active objectives may vary by codebook level. The approach may be deployed as a drop-in replacement for scalar top-B beam pruning without model retraining.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Pareto Beam Search", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10639