Technology method · wiki:deep

MRL training

Matryoshka Representation Learning trains O(log d) nested prefix sizes so each cutoff is a valid representation — the reason OpenAI dimensions and mxbai truncate work.

MRL training flow

MRL (Kusupati et al., NeurIPS 2022 lineage / RAIVNLab) modifies representation learning so a single d-dimensional vector encodes usable prefixes at chosen sizes M (often powers of two up to d). Losses are applied at each mM; intermediate lengths often interpolate. MRL–E ties classifier weights across prefixes for efficiency.

Product APIs do not expose the training loop; they expose the consequence: Matryoshka truncation. OpenAI text-embedding-3-* and mxbai-embed-large are deployment surfaces of that idea (vendor training details are partial — do not invent the exact nesting set).

Primary research artifacts for this cell: the MRL paper + RAIVNLab/MRL GitHub (loop wave). Sentence-Transformers documents Matryoshka loss helpers for OSS fine-tunes.

See also

In this taxonomy

_catalog/taxonomy/tech-mrl-training.md