Microsoft's EigenLI: training-free spectral compression of ColBERT-style representations

lateinteraction · x · 2026-09-10

Microsoft researchers present EigenLI, a training-free spectral method that compresses ColBERT-style multi-vector representations. The key observation is that multi-vector representations lie in low-rank, seemingly document-specific subspaces, and the method exploits this for practical compression. It outperforms clustering pooling and MUVERA single-vector baselines, while raising new questions about the geometry of multi-vector model spaces.

Related event: Microsoft's EigenLI Compresses ColBERT Representations Without Training(2 posts)→

Original post →

More from Research

Research channel →