At the Korea Computer Congress 2026 (KCC 2026), held over three days from 24 to 26 June, the paper 'Domain-Based Dimension Analysis for Computational Efficiency in Multi-Vector Retrieval', presented by master's student Cho Hyun-ji (supervisor: Park Sung-yong), received the Best Paper Award in the high-performance computing field.

Multi-vector retrieval represents queries and documents as multiple token vectors and performs MaxSim operations between query tokens and document tokens. The amount of MaxSim computation grows in proportion to the number of query tokens, the number of document tokens and the embedding dimensionality, but existing optimization research has focused chiefly on reducing the number of token vectors and has not sufficiently explored the potential for dimension-level computational efficiency. This research designs a dimension importance model to find the important dimensions that contribute most to forming the similarity score in MaxSim computation. The model jointly reflects each dimension's contribution to the similarity score, the consistency of that contribution, and its influence on document token selection.

Comparing the distribution of dimension importance across domains revealed regions of dimensions that are commonly important in several domains. This means that dimension importance is not tied to individual queries alone but can be shared within the same domain, showing the potential for computational efficiency through domain-level common dimension profiling.