- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes AC-FDE, a method that advances multi-vector approaches such as MUVERA. MUVERA works by turning multi-vector representations into single fixed-dimensional encodings (FDEs) by repeated random SimHash bucketing and aggregation. AC-FDE targets MUVERA's efficiency-accuracy trade-off: a good MaxSim approximation needs many repetitions, which makes the FDE vectors very high-dimensional. With AC-FDE, the authors propose adding two offline steps, both of which are fitted to a small random sample of documents (10% by default). First, it selects repetitions without considering queries. Second, it compresses the encodings with PCA run separately on each repetition. At query time, the stored repetitions and projections are applied to the query FDE, followed by HNSW search and MaxSim reranking. Experiments on MS MARCO, SciDocs, FiQA, and NQ report up to 4.3x speedup over MUVERA, LEMUR, and multi-vector indexing baselines at equal recall, better pre-reranking candidate recall than LEMUR, and small offline overhead over MUVERA.
Critical strengths
- Interesting idea that addresses a real practical problem.
- AC-FDE is a simple and cheap method.
Critical weaknesses
The paper proposes an interesting idea. However, several problems affect the current submission, which is IMHO not ready for publication in its current shape. Here are some detailed points
-
The experimental evaluation can be significantly improved. Several points on this: 1) Table 1 reports only the comparison between AC-FDE and LEMUR. Why not add MUVERA too? You added it as an evaluated method in Section 8.1. 2) The use of 1-Recall@N, although the authors provide a pointer to prior work, is not explained. Why is this metric used in this kind of evaluation? 3) Ground truth is not defined in the main text. 4) Table 1 uses HNSW (efSearch=500). The analysis mixes encoding error with approximate-search error. 5) MUVERA is run with PQ, and the multi-vector index baselines (PLAID, EMVB, IGP, WARP) appear with only a few operating points in Fig. 2. It is not clear how they were tuned to their own Pareto frontiers. Please describe how the multi-vector index baselines (PLAID, EMVB, IGP, WARP, DESSERT, GEM) are configured and tuned. 6) A recent SOTA multi-vector index is not tested nor discussed [1]. I suggest authors to add and compare to it.
-
In terms of writing and presentation, the main paper (first nine pages) is too dense and difficult to read. Six theorems are packed into Sections 2-5, with little intuition provided. Figures 2-4 are hard to read at their printed size.
In its current form, the paper is not ready for publication. I suggest that the authors move the key evidence for repetition selection into the main text, shorten the theory in favor of intuition, fully specify the experimental setup, and add the missing comparisons (especially MUVERA).
Minor things:
- Algorithm 1: "Equation equation 6"
- "Validataion", end of page 33
[1] Martinico et al. Efficient Multivector Retrieval with Token-Aware Clustering and Hierarchical Indexing. ACM SIGIR 2026.
Questions for authors
see my weaknesses
Thank you for the thorough rebuttal and additional experiments. I keep my original rating.