TopoGR preserves semantic-ID topology for generative recommendation with Hamming geometry

_reachsumit · x · 2026-07-29

TopoGR targets generative recommendation by addressing a mismatch between semantic-ID tokenization and generation.
The paper argues that existing methods treat semantic IDs as independent symbols, losing neighborhood structure and making similarity depend too much on exact SID overlap.
It proposes Bit-decomposable Semantic IDs (Binary SID), which are deterministic to standard integer IDs but expose explicit Hamming geometry.
The framework uses this topology in three stages: Hamming-preserving input features, Hamming soft targets, and Hamming-consistent reranking at inference.

More from coding & agent