CrossQ: conditional quantization for late interaction

A transparent toy reproduction of the mechanism, not a paper-scale benchmark.

Claim 1 · index pressure

8 × 32

Each synthetic document stores eight token vectors of dimension 32. Full precision uses 32 bits per scalar, illustrating why multi-vector indices grow quickly.

Claim 2 · conditioning

2-bit + context

Document mean context is computed once at indexing time; token residuals are quantized relative to that context and reconstructed before MaxSim scoring.

Claim 3 · result

9.85×

Our toy proxy reaches 9.85× conditional index compression. The released checkout has no paper code, weights, or benchmark data, so 61× and 2.3% are not verified.

Reproducibility status

Local run: deterministic NumPy script, 120 documents, 48 queries. HF GPU Job: blocked before launch by account credit (HTTP 402). See the linked logbook pages for commands and outputs.

OpenReview paper · Posterly