SciFACE Framework Enables Controllable Diversity in Paper Recommendations Through Facet-Aware Reranking

A novel reranking system known as SciFACE (Scientific Faceted Cross-Encoder) has been developed to overcome the shortcomings of existing paper recommendation tools by focusing on two distinct aspects: Background (the problem addressed) and Method (the solution approach). It was trained on 5,891 authentic seed-candidate paper pairs, which were labeled by GPT-4o-mini based on specific facet criteria, with these labels corroborated by human assessments. On the CSFCube benchmark, SciFACE recorded an NDCG@20 score of 70.63 for Background recommendations, surpassing SPECTER by 5.9 points. For Method recommendations, it achieved 49.06 NDCG@20, outpacing SPECTER by 31.1 points. Compared to FaBLE, which lacked citation pre-training, SciFACE enhanced Method NDCG@20 by 4.1 points while utilizing far fewer training examples—5,891 labeled pairs instead of 40,000 synthetic augmentations. This research, identified by arXiv as 2604.16329v1, highlights the efficiency of high-quality grounded facet labels over synthetic ones. The framework allows users to indicate the reasons for paper similarities rather than receiving a singular mixed similarity score.