Microsoft's EigenLI Compresses ColBERT Representations, Beats MUVERA Baseline

_reachsumit · x · 2026-09-09

Microsoft researchers present EigenLI, a training-free spectral method that compresses ColBERT-style late-interaction representations. Key insight: document token embeddings concentrate in a low-dimensional subspace that preserves most retrieval signal, so each document's dominant eigendirections can build a reduced interaction representation. k-EigenLI with k≤32 outperforms k-means and Ward clustering pooling on ColBERTv2 and AnswerAI-ColBERT-small, though GTE-ModernColBERT favors clustering at k=32. The same construction also yields EigenLI-SV, an ANN-compatible single-vector representation that consistently beats MUVERA-like surrogates across datasets and models.

Original post →

More from Research

Research channel →