Controlled test: multi-vector beats dense on 9 of 13 NanoBEIR benchmarks

tomaarsen · x · 2026-08-18

Does late interaction actually help? LightOn trained LateOn (multi-vector) and DenseOn (dense) on the same data with the same 149M ModernBERT backbone, differing only in pooling. Multi-vector wins 9 of 13 NanoBEIR datasets — 0.6868 vs 0.6764 mean nDCG@10 — with the same gap on full BEIR. In v6.0, every checkpoint format (PyLate, Stanford-NLP ColBERT, ColPali-style VLMs, or a bare backbone with a fresh projection) loads through one class, with prefixes, query expansion and the punctuation skiplist coming from the saved config.

Related event: Sentence Transformers v6.0 ships with first-class late interaction models(33 posts)→

Original post →

More from coding & agent

coding & agent channel →