We study the problem of achieving high efficiency in iterative combinatorial auctions (ICAs). ICAs are a kind of combinatorial auction where the auctioneer interacts with bidders to gather their valuation information using a limited number of queries, aiming for efficient allocation. Preference elicitation, a process that incrementally asks bidders to value bundles while refining the outcome allocation, is a commonly used technique in ICAs. Recently, the integration of machine learning (ML) into ICAs has significantly improved preference elicitation. This approach employs ML models that match the number of bidders, estimating each bidder’s valuation functions based on their reported valuations. However, most current studies train a separate model for each bidder, which can be inefficient when there are numerous bidders with similar valuation functions and a limited number of available queries. In this study, we introduce a multi-task learning method to learn valuation functions more efficiently. Specifically, we propose to share model parameters during training to grasp the intrinsic relationships between valuations. We assess the performance of our method using a spectrum auction simulator. The findings demonstrate that our method achieves higher efficiency than existing methods, especially in scenarios with many bidders and items but a limited number of queries.