Search papers, labs, and topics across Lattice.
This paper analyzes the internal dynamics of Late Interaction retrieval models, focusing on length biases in multi-vector scoring and the utility of similarity scores beyond the top-ranked document token. They find that while causal Late Interaction models exhibit a theoretical length bias, bidirectional models can also be affected in extreme cases. Their analysis on NanoBEIR also indicates that the MaxSim operator effectively captures relevant token-level similarity, as no significant similarity trend exists beyond the top-1 document token.
Bidirectional Late Interaction models, despite their architecture, can still fall prey to length biases, challenging assumptions about their robustness.
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.
