Skip to main content
preprint arXiv preprint 2026

Hubness, Not Anisotropy, Drives Cross-Lingual Retrieval Asymmetry in Multilingual Embedding Models

Adib Sakhawat, Fardeen Sadab, Atik Shahriar

Recommends cosine similarity with CSLS as the default retrieval metric for multilingual embedding pipelines.

BibTeX

@misc{sakhawat2026hubness,
  title         = {Hubness, Not Anisotropy, Drives Cross-Lingual Retrieval Asymmetry in Multilingual Embedding Models},
  author        = {Sakhawat, Adib and Sadab, Fardeen and Shahriar, Atik},
  year          = {2026},
  eprint        = {2605.26575},
  archivePrefix = {arXiv}
}

Description

The question

Multilingual embedding models are deployed on an assumption of symmetry: if a query in language A retrieves its translation in language B, the reverse should also hold. In practice it often doesn't. Why?

The setup

  • A parallel corpus of 6,518 idiomatic and proverbial expressions in English, Bangla, Hindi and Arabic.
  • Five production embedding models: Gemini, Mistral, OpenAI-L, OpenAI-S and Qwen.
  • Failure formalised as a deficit in mutual nearest-neighbour reciprocity.
  • Five pre-registered experiments, each with falsification conditions specified in advance.

The single claim under test: among the known geometric pathologies of multilingual spaces, hubness (a few vectors being nearest neighbours to far too many others), not anisotropy, centroid drift or magnitude, is the dominant cause.

Results

Joint regression of reciprocity on geometric predictors:

Predictor Partial R² Dominance share p
Hub mass 0.302 49.5% 0.001
Dimension 0.150 29.4% 0.012
Centroid drift 0.003 8.7% 0.668
Byte ratio 0.072 7.7% 0.065
Anisotropy 0.003 4.7% 0.677

Key findings

  • Hubness dominates. Hub mass carries 49.5% of the explained variance, 1.68 times the next predictor, while anisotropy contributes almost nothing.
  • Fix the metric, not the vectors. A hub-aware score correction (CSLS) closes 63.5% of the worst-to-best reciprocity gap, and its effect is 130 times larger than surgically ablating hub vectors. Hubness is a pathology of the similarity metric, not of individual points.
  • The paradox, resolved. Anisotropy and hubness are statistically dissociable, which settles the long-standing anisotropy–hubness question.

Recommendation

Replace plain cosine similarity with CSLS as the default retrieval metric in multilingual embedding pipelines.

Resources