Late-Interaction InfoNCE: A Novel Contrastive Loss Function with Combined Similarity Metrics

Multi-hop reasoning presents significant challenges in natural language processing, particularly due to asymmetric retrieval where evidence within a reasoning chain varies greatly in retrieval difficulty. Implicit evidence, which lacks explicit retrieval identifiers, is often overshadowed by more readily retrievable explicit evidence, leading models to favor superficial associations over deeper semantic reasoning. To address this, we propose Late-Interaction InfoNCE, a novel contrastive loss function that integrates both dot product and Hadamard product interactions to enhance the representation of semantic relationships between queries and evidence samples. This method improves the model's ability to capture fine-grained feature couplings and global semantic alignments simultaneously. Additionally, we introduce a false negative masking strategy to reduce interference from semantically related negative samples. Experiments on the EX-FEVER dataset demonstrate that our approach achieves state-of-the-art performance across multiple retrieval metrics, including Recall and MRR, validating its effectiveness in complex asymmetric retrieval scenarios.