CrossQ: Task-Aligned Quantization for Late Interaction Retrieval

_reachsumit · x · 2026-08-21

The paper presents CrossQ, a compression method for ColBERT-style late interaction retrieval. Unlike standard quantization minimizing reconstruction error, CrossQ conditions token codes on document context and trains with ranking-aligned objectives. It achieves a 64x storage reduction at 4 bits/token while retaining approximately 98% of full-precision performance.

Original post →

More from Research

Research channel →