Abstract
OxygenREC is a next-generation generative recommendation (GR) framework from JD.com designed to bridge the gap between deep LLM-based reasoning and the strict latency requirements of industrial e-commerce. By adopting a "Fast-Slow Thinking" architecture, it leverages near-line Large Language Models (LLMs) to generate deductive reasoning instructions that condition a high-efficiency real-time backbone. This framework achieves a "train-once-deploy-everywhere" paradigm, significantly improving performance across multiple shopping scenarios while maintaining industrial-grade latency.
Introduction
Traditional recommendation systems typically rely on a cascaded pipeline consisting of retrieval, pre-ranking, and ranking stages. While effective, this architecture often suffers from fragmented optimization and misaligned goals across stages. Generative Recommendation (GR) has emerged as a promising alternative, treating recommendation as an end-to-end sequence generation task. However, existing GR models face two primary challenges: a "Reasoning Gap" where models lack deductive capabilities, and a "Scalability vs. Latency Trade-off" due to the high computational cost of LLMs.
Methodology
OxygenREC introduces a novel "Fast-Slow Thinking" architecture inspired by cognitive psychology to decouple complex reasoning from real-time execution.
Fast-Slow Thinking Architecture
The system is bifurcated into two components:
- Slow Thinking (Near-line): An LLM-based pipeline that synthesizes Contextual Reasoning Instructions. These instructions capture deep user intent (e.g., "User is looking for professional photography gear") using world knowledge.
- Fast Thinking (Online): A high-throughput encoder-decoder backbone that consumes these instructions to generate item recommendations in real-time.
Key Components
- Instruction-Guided Retrieval (IGR): Uses reasoning instructions as queries to filter user history, ensuring the model focuses on intent-relevant actions.
- Soft Adaptive Group Clip Policy Optimization (SA-GCPO): A reinforcement learning algorithm that utilizes a smooth, sigmoid-based gate to stabilize training and better distinguish between positive and negative user feedback.
- Unified Reward Model: A single transformer-based model acts as a reward mapping service, allowing the backbone to serve diverse business objectives (e.g., GMV, diversity) across different app scenarios without requiring scenario-specific models.
Experimental Results
The framework was evaluated using massive industrial datasets from JD.com, covering search logs and multiple recommendation scenarios such as Homepage and Channel Feeds.
Business Impact
Deployment in JD.com's production environment yielded significant gains. In the "Channel Feeds" scenario, OxygenREC achieved an 8.03% increase in Order Volume and a 7.89% increase in Conversion Rate (UCTCVR).
System Efficiency
The system maintains high efficiency, with the training framework reaching 40% Model FLOPs Utilization (MFU). The custom inference engine, xGR, handles large-beam decoding (sizes 256-512) within a strict 50-80ms latency window, proving its viability for high-traffic industrial applications.
Conclusion and Future Work
OxygenREC demonstrates that integrating LLM-level reasoning into a generative framework can solve complex intent gaps while meeting industrial latency requirements. Future research will focus on addressing the decoding bottleneck through Non-Autoregressive (NAR) parallel generation and enhancing multi-scenario user trajectory modeling to capture long-term decision-making processes.
Summary updated