Snap Paper: Fixing Lost Semantics in LLM-Based Recommenders

_reachsumit · x · 2026-08-11

Shared a paper from the Snap team titled "Preserving Item Semantics for Free".

The paper reveals that semantic ID tokens in LLM-based recommenders end up encoding popularity rather than semantics. The team fixes this issue with a free centroid initialization method.

Original post →

More from Research

Research channel →