Meta's LIGE-GR upgrades itemwise recommenders to listwise LLM-style generative ranking

_reachsumit · x · 2026-09-17

A large Meta team published LIGE-GR on arXiv, addressing how to absorb the LLM paradigm's sequence-level generation and optimization into mature industrial recommender systems.

The approach adds a lightweight context-aware module plus an RL-based decoder to existing itemwise recommenders, smoothly upgrading them to listwise, LLM-style generative ranking. The paper (62 authors) tackles two core challenges: incorporating sequence-level generation into recommendation, and retrofitting already-mature production systems.

Original post →

More from Research

Research channel →