FLINT: Fixed-budget Late Interaction
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.
ProofPaper ↗
Key points
- Multi-vector retrieval models such as ColBERT often outperform single-vector models of comparable scale by matching queries and documents at the token level.
- 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.
- 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.
Sources (1)
- [1]FLINT: Fixed-budget Late InteractionarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 08:46 PM
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.
Multi-vector retrieval models such as ColBERT often outperform single-vector models of comparable scale by matching queries and documents at the token level.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026BERT Is Already a Denoiser: Cheaply Unlocking Iterative Generation in Masked Language Models
- Oct 19, 2026SNO-PI: Label-Scarce SNOMED CT Entity Linking with Pre-trained Encoders and Curated Terminology Knowledge
- Oct 19, 2026ClueFlow: Learning Long-Horizon Search through Attributed Research-State Transitions
- Oct 19, 2026OpenSaC: A Python Library for Search as Code in Deep Research Agents