Snap uses minimum set cover to optimize decoding tries for generative recommenders

_reachsumit · x · 2026-07-21

## Snap’s generative recommendation work optimizes decoding tries with co-designed text IDs Snap presents **“Beyond Fixed Depths and Widths: Optimizing Textual Decoding Tries in LLM-based Generative Recommendation.”** The paper describes a co-design between **textual term IDs** and the **decoding trie**. - It uses **minimum set cover** to choose term IDs that better fit the trie structure. - The goal is to improve **LLM-based generative recommenders** by making decoding more efficient and better aligned with the recommendation vocabulary. - This is a systems + modeling optimization paper, not just a product announcement.

Original post →

More from Apps

Apps channel →