Late-interaction retrieval models rely on hard maximum similarity (MaxSim) to aggregate token-level similarities. Although effective, this winner-take-all pooling rule may structurally bias training dynamics. We provide a mechanistic study of gradient routing and robustness in MaxSim-based retrieval. In a controlled synthetic environment with in-batch contrastive training, we demonstrate that MaxSim induces significantly higher patch-level gradient concentration than smoother alternatives such as Top-k pooling and softmax aggregation. While sparse routing can improve early discrimination, it also increases sensitivity to document length: as the number of document patches grows, MaxSim degrades more sharply than mild smoothing variants. We corroborate these findings on a real-world multi-vector retrieval benchmark, where controlled document-length sweeps reveal similar brittleness under hard max pooling. Together, our results isolate pooling-induced gradient concentration as a structural property of late-interaction retrieval and highlight a sparsity-robustness tradeoff. These findings motivate principled alternatives to hard max pooling in multi-vector retrieval systems.
Figures & tables
Figure 1: Synthetic training dynamics and document-length sweep. (A) Patch-level gradient concentration (Gini) vs. training. (B) Retrieval quality (Recall@1) vs. training. (C) Retrieval quality vs. document length M under fixed queries.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 2: Token–patch similarity heatmaps for a representative ColQwen2.5 example from ViDoRe biomedical retrieval. Left (Baseline, K=0 ): Token maxima align with semantically relevant document patches. Middle (Hard-negative injection, K=100 ): Injected distractor patches redirect the majority of token-wise argmax selections into the injected region, resulting in ∼ 83% token hijacking. Right (Gaussian control, K=100 ): Random noise produces minimal routing changes and low hijack rates. Black dots denote argmax positions on original patches; red dots denote argmax positions on injected distractor patches. The injected region is highlighted.
Late interaction retrieval matches query and document token vectors after encoding to preserve fine-grained relevance while supporting efficient search.
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.
Multi-vector retrieval models such as ColBERT often outperform single-vector models of comparable scale by matching queries and documents at the token level. However, retaining a vector for every token substantially increases storage requirements and slows retrieval compared to single-vector bi-encoders, limiting their scalability. We introduce FLINT (Fixed-budget Late INTeraction), an architecture that uses attention-based pooling to represent each query and document with a fixed number of embedding vectors. Unlike architectures that pool only document representations while retaining token-level queries, FLINT applies pooling symmetrically to both sides, limiting document storage and the number of vector comparisons needed during retrieval. FLINT achieves comparable ranking effectiveness to the closest baselines on BEIR and outperforms them on several benchmarks with longer and more complex queries, including BrowseComp+, CRUMB, and TREC Tip-of-the-Tongue 2025. We demonstrate that FLINT's effectiveness-efficiency benefits extend across encoder architectures, testing it on two separate backbones (BERT and ModernBERT). On long-query corpora, FLINT delivers 1.5-2.8 times faster end-to-end retrieval than ConstBERT, the closest baseline. FLINT also reduces index size by 5.3-9.3 times relative to token-level late-interaction models such as ColBERTv2.
Himank Handa, Negar Arabzadeh, Matei Zaharia
University of Illinois Urbana-Champaign · University of California, Berkeley
Late-interaction visual document retrievers preserve fine-grained page evidence by storing many token embeddings per page, but the resulting storage and query-time interaction costs make large-scale deployment expensive. Pooling document tokens before indexing offers a natural remedy, yet static pooling must decide which visual evidence to preserve before the query is known. We study an alternative: a heavily compressed hot-path index generates candidates, after which query-aware token budgeting operates on the original token sets of the shortlisted pages. We formulate this stage-two selection as a budgeted MaxSim coverage problem, show that a clipped version is monotone submodular, and compare coverage-only, cluster-guided, token-wise, and marginal-gain policies. On ten ViDoRe tasks with ColModernVBERT, direct static pooling reduces macro normalized discounted cumulative gain at rank five from 0.6309 without compression to 0.4738 at a thirty-two-fold pool factor. Under the same candidate-generation regime and a pool-factor-eight-equivalent reranking budget, token top-k recovers 93.93 percent of the full-token score, while greedy marginal-gain selection recovers 98.39 percent. Held-out and leave-one-dataset-out evaluations yield positive greedy improvements over token top-k on every dataset. The latency analysis reveals two useful operating points: token top-k for interactive retrieval and the naive greedy implementation as a quality upper envelope. Together, these results show that late-interaction visual retrieval benefits from query-aware allocation rather than query-agnostic pooling alone.
PS Rishi, Rajeev Ranjan Dwivedi, Vinod K Kurmi
Indian Institute of Science Education and Research Bhopal (IISER Bhopal), India