Stefan Josef

The paper proposes lightweight fine-tuning as an alternative to pooling-aware training for reducing vector count in late interaction models, achieving broad gains and high compression rates.