The paper "Mixture of Experts approaches in Dense Retrieval Tasks" by Effrosyni Sokli, Pranav Kasela, Georgios Peikos and Gabriella Pasi (IKR3 Lab) has received the Best Paper Award at the 2025 IEEE/WIC International Conference on Web Intelligence and Intelligent Agent Technology (WI-IAT 2025), that took place in London, United Kingdom, from the 15th until the 18th of November 2025.
In this year's edition of WI-IAT, there were 274 submissions and an overall acceptance rate of 32.12%. This contribution stood out and was selected for the award among a total of 68 long papers that were published in the proceedings of the conference.
The paper introduces SB-MoE: a framework that enhances neural Information Retrieval (IR) models by leveraging Mixture-of-Experts (MoE). The authors investigate different Mixture-of-Experts settings integrated with state-of-the-art Transformer-based dense retrievers and conduct various analyses supported by empirical evaluations and visualizations to assess the impact of MoE approaches in dense retrieval tasks.