A July 2, 2026 brief on a proposal to address vocabulary gaps in learned sparse retrieval with advanced encoders.
Published on July 2, 2026, the brief attributes advanced encoders' lower performance in learned sparse retrieval, compared with BERT-base, to a vocabulary gap. It says modern tokenizers map semantic units to redundant surface forms; the paper proposes Vocabulary Transfer to address this incompatibility.
The proposal may interest engineers evaluating sparse retrieval methods. The brief gives no quantitative results, implementation details, or experimental conditions. To verify the claim's scope, consult the original paper and check its method, evaluation setup, and comparative results; those details are not in the supplied material.