LQ-Rec: Late Quantization Makes Longer Semantic IDs Improve Recommendation Ranking

_reachsumit · x · 2026-11-20

LQ-Rec (Koide, Wang, McAuley) tackles a structural scaling limit in generative retrieval for sequential recommendation: in autoregressive GR, longer semantic IDs reduce quantization error but degrade ranking and cost quadratic memory.

The fix is Late Quantization—confine discretization to the output side. The encoder and scoring operate in continuous space; items are stored as semantic IDs from an end-to-end residual quantizer, enabling an additive error decomposition and a pairwise ordering theorem that bound quantization-induced ranking drop.

On four public benchmarks: ranking improves monotonically with semantic-ID length and beats baselines; one model supports anytime compression (tune quality vs. inference cost without retraining); and it generalizes better on memorization/generalization partitions.

Original post →

More from Research

Research channel →