Pith. sign in

REVIEW 4 cited by

Multi-Vector Retrieval as Sparse Alignment

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2211.01267 v1 pith:37R33OJP submitted 2022-11-02 cs.CL cs.IR

Multi-Vector Retrieval as Sparse Alignment

classification cs.CL cs.IR
keywords retrievalalignmentsdocumentmulti-vectorsaliencesunaryperformancetoken
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multi-vector retrieval models improve over single-vector dual encoders on many information retrieval tasks. In this paper, we cast the multi-vector retrieval problem as sparse alignment between query and document tokens. We propose AligneR, a novel multi-vector retrieval model that learns sparsified pairwise alignments between query and document tokens (e.g. `dog' vs. `puppy') and per-token unary saliences reflecting their relative importance for retrieval. We show that controlling the sparsity of pairwise token alignments often brings significant performance gains. While most factoid questions focusing on a specific part of a document require a smaller number of alignments, others requiring a broader understanding of a document favor a larger number of alignments. Unary saliences, on the other hand, decide whether a token ever needs to be aligned with others for retrieval (e.g. `kind' from `kind of currency is used in new zealand}'). With sparsified unary saliences, we are able to prune a large number of query and document token vectors and improve the efficiency of multi-vector retrieval. We learn the sparse unary saliences with entropy-regularized linear programming, which outperforms other methods to achieve sparsity. In a zero-shot setting, AligneR scores 51.1 points nDCG@10, achieving a new retriever-only state-of-the-art on 13 tasks in the BEIR benchmark. In addition, adapting pairwise alignments with a few examples (<= 8) further improves the performance up to 15.7 points nDCG@10 for argument retrieval tasks. The unary saliences of AligneR helps us to keep only 20% of the document token representations with minimal performance loss. We further show that our model often produces interpretable alignments and significantly improves its performance when initialized from larger language models.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Near-Optimal Dimension Lower Bounds for Single-Vector Embeddings of Maximum Inner Product Similarity

    cs.DS 2026-07 accept novelty 8.0

    Single-vector embeddings of MAX-IP similarity require dimension at least m^{c/ε^{2−2δ}} for any δ>0, nearly matching the m^{O(1/ε²)} upper bound.

  2. Multi-Vector Embeddings are Provably More Expressive than Single Vector Embeddings

    cs.DS 2026-06 unverdicted novelty 8.0

    Proves that for bounded n, there exist MV embeddings with m vectors each whose Chamfer matrix requires single-vector dimension D = (ε² m)^Ω(1/ε) to approximate within ε, separating MV from SV expressiveness.

  3. Visual Late Chunking: An Empirical Study of Contextual Chunking for Efficient Visual Document Retrieval

    cs.CV 2026-04 unverdicted novelty 7.0

    ColChunk adaptively chunks visual document patches into contextual multi-vectors via clustering, cutting storage by over 90% while raising average nDCG@5 by 9 points.

  4. A Voronoi Cell Formulation for Principled Token Pruning in Late-Interaction Retrieval Models

    cs.IR 2026-03 unverdicted novelty 6.0

    A Voronoi cell estimation framework in embedding space enables principled token pruning for late-interaction models, reducing index size while retaining retrieval quality.