🤖 AI Summary
This study systematically investigates length bias in late-interaction retrieval models and the characteristics of similarity distributions beyond the MaxSim pooling mechanism. Through fine-grained analysis of prominent multi-vector late-interaction models on the NanoBEIR benchmark, the work provides the first empirical evidence that both causal and bidirectional architectures are susceptible to length bias. Furthermore, it reveals that suboptimal matching signals—those beyond the top-1 similar token—contribute negligible retrieval gains, thereby validating the sufficiency of MaxSim in capturing effective semantic matches. These findings not only confirm the practical existence of length bias but also offer theoretical support, from the perspective of similarity distribution, for the widespread effectiveness of the MaxSim operator in dense retrieval systems.
📝 Abstract
While Late Interaction models exhibit strong retrieval performance, many of their underlying dynamics remain understudied, potentially hiding performance bottlenecks. In this work, we focus on two topics in Late Interaction retrieval: a length bias that arises when using multi-vector scoring, and the similarity distribution beyond the best scores pooled by the MaxSim operator. We analyze these behaviors for state-of-the-art models on the NanoBEIR benchmark. Results show that while the theoretical length bias of causal Late Interaction models holds in practice, bi-directional models can also suffer from it in extreme cases. We also note that no significant similarity trend lies beyond the top-1 document token, validating that the MaxSim operator efficiently exploits the token-level similarity scores.