MRVQ paper: one elastic vector search index cuts memory up to 22x vs per-rate quantizers
_reachsumit · x · 2026-10-05
A new arXiv paper introduces Matryoshka Residual Vector Quantization (MRVQ), a post-hoc residual quantizer for frozen embeddings that serves dense retrieval with a single resident index elastic in both dimension and bit rate.
- How it works: the max-rate code can be truncated two ways — dropping residual stages lowers the rate, dropping embedding coordinates lowers the dimension — so one artifact serves every (dimension, rate) pair, avoiding separately trained quantizers per rate.
- Results: across FiQA and NFCorpus, four embedding families, and {4, 8, 16}-byte codes, MRVQ is the lowest-RAM design: 17.8–22.0x less memory than three separately trained QINCo2 indices and 1.89–2.02x less than a lean shared-model steelman; it beats PQ, OPQ, and AdANNS-OPQ at matched code size.
- Trade-off: per-rate QINCo2 is 0.026–0.107 nDCG@10 better on FiQA, so MRVQ is a low-memory operating point, not a universal quality winner.
- Extras: a low-build-cost PCA-scalar design matching RaBitQ quality while fitting 420x faster at the median, plus two reported negative results.
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23