The Pulse
SpecMRL Raises MTEB Scores With Half-Size Embeddings
SpecMRL reports a higher English MTEB score than standard MRL at half the embedding dimensions on Qwen3-0.6B, though its full-dimensional score is slightly lower. The paper also reports faster retrieval for half-dimensional vectors in a spe
AI.info Team · Written with AI · fact-checked ·
A stronger score at half the size, but not at full size
A new embedding method reports a better English MTEB score with half the dimensions of its standard baseline, while falling just short when both methods use the full vector. The researchers behind SpecMRL say their approach improves Matryoshka Representation Learning (MRL) at every reduced size they tested, with larger gains under stronger compression. The work appeared as a preprint on arXiv on October 19, 2026.
On Qwen3-0.6B, the paper’s average MTEB score rises from 60.54 with standard MRL to 61.61 with SpecMRL at half the original 1,024 dimensions. At full size, the results are 60.97 for MRL and 60.94 for SpecMRL. At one-sixteenth size, SpecMRL scores 58.53 against MRL’s 54.73, a 3.80-point gap. The headline claim therefore rests on the compressed settings, not a uniform improvement at every dimension.
“Notably, SpecMRL extends the performance frontier of MRL, surpassing full-dimensional MRL with only half the embedding dimensions.”
Zhongxin Chen, Songhao Wu, Heng Cui, Cong Li, Xiao Zhou and Rui Yan, the paper’s authors, write
Reordering the dimensions before truncation
MRL trains embeddings so that leading prefixes can serve as smaller representations. SpecMRL changes the coordinate system those prefixes use: it derives an ordered spectral basis from a learned projection matrix, then reads the embedding in that basis. The idea is to put more useful feature directions earlier, so cutting off the rest loses less information.
The authors describe the method as a modification to the existing MRL training setup rather than a replacement for its objective. They approximate the spectral basis with a power-iteration procedure and report throughput comparable to standard MRL during training. Once training ends, the basis is fixed; encoding uses a projection before the vector is stored or searched.
A speed test on one million MS MARCO passages
The paper separately measures retrieval using one million MS MARCO passages, with 8,192 passages selected as queries and the rest used as the database. The test uses FAISS exact inner-product search in FP32 on a single NVIDIA A100 PCIe GPU, with the indexes held in GPU memory. At a query batch of 1,024, half-dimensional vectors reach 14,605 queries per second, compared with 7,859 for full-dimensional vectors—a reported 1.86× speedup.
That speed figure comes from a particular GPU, search method and batch size, not a general deployment benchmark. The paper reports that encoding 100,000 passages takes 313.99 seconds with SpecMRL versus 312.85 seconds with MRL, a 0.36% increase that includes the additional projection. For batch size one, the half-dimensional retrieval speedup is smaller: 1.26×.
What the evaluation does—and does not—show
The main results average scores across the English MTEB benchmark’s 56 datasets and seven task categories, including retrieval, clustering and semantic similarity. The authors also report a 4B-parameter Qwen3 evaluation, where SpecMRL’s score exceeds MRL’s at several reduced dimensions; at 2,560 dimensions, however, it scores 64.29 against MRL’s 64.71. The reported findings are preprint results, and the authors’ evidence is benchmark-based rather than a comparison across live search services.
For teams weighing embedding storage against search throughput, the paper reports a higher average MTEB score than full-size MRL in its half-size Qwen3-0.6B setting. Separately, its retrieval test reports a 1.86× throughput improvement for half-dimensional vectors at batch size 1,024 under the stated A100 conditions.