GR2 re-ranker boosts recommendations with RL-trained LLM reasoning

Large language models have struggled to make an impact on industrial recommendation systems, especially at the critical re-ranking stage that directly determines what users see. A new technical report introduces GR2 (Generative Reasoning Re-Ranker), an end-to-end framework that finally bridges this gap. The team—likely from a major tech company based on the scale and author list—tackles three key obstacles: LLMs rarely target re-ranking; they underutilize reinforcement learning; and they can't handle billions of non-semantic item IDs. GR2 solves this by mid-training on semantic tokens (≥99% unique), distilling reasoning traces from a stronger teacher via targeted prompting and rejection sampling, and applying RL with verifiable rewards purpose-built for re-ranking. A novel context compressor reduces training costs, while On-Policy Distillation replaces collapsing supervised fine-tuning at scale.

On real industrial traffic, GR2 delivers substantial gains: +18.7% Recall@1, +7.1% Recall@3, and +9.6% NDCG@3 over legacy baselines. The paper highlights that reward design is critical—naive rewards cause LLMs to hack the system by preserving input order or exploiting position bias. The team introduces conditional verifiable rewards to prevent such gaming. This work could finally make generative models practical for the most user-facing part of recommendation systems, powering better search results, carousels, and grids for billions of users. The report also details reasoning distillation for low-latency serving, making the approach production-ready.