OxygenREC-v2: Internalizing Discrimination into Generative Recommendation

2026-07-27

AI Summary

Existing generative recommendation approaches struggle to effectively integrate multi-behavior signals—such as clicks, add-to-cart actions, and purchases—often suffering from performance limitations due to conflicting objectives or misaligned rewards. This work proposes IDGR, a novel framework that internalizes discriminative capabilities directly into the generative process by guiding generation with behavior-specific instructions and leveraging future user interactions as privileged knowledge for reward-free policy optimization, thereby circumventing explicit multi-objective trade-offs and external reward models. Built upon a sparse Mixture-of-Experts architecture with 3B parameters and 1B active units, IDGR incorporates behavior-conditioned pretraining and entropy-aware trajectory self-distillation. Deployed on JD.com’s large-scale e-commerce platform, it achieves 1.6–4.4% gains in click-through conversion rate and 2.8–6.8% improvements in GMV over OxygenREC-v1 in online A/B tests.

Abstract

Generative recommendation unifies retrieval and ranking within a single model by autoregressively decoding semantic identifier (SID) sequences. Yet reliably incorporating behavior signals from clicks, cart additions, and orders remains challenging. Existing approaches either jointly optimize generative and discriminative objectives, requiring delicate trade-offs, or use a separate ranker as a post-hoc reinforcement-learning reward, risking out-of-distribution scoring and reward misalignment. We propose OxygenREC-v2, a generative recommender that Internalizes Discrimination into Generative Recommendation (IDGR). Rather than adding a separate discriminative objective, OxygenREC-v2 uses logged behavior to condition generation and supervise training. During pre-training, a behavior instruction conditions generation on the target behavior. During post-training, future interaction behaviors are exploited as privileged knowledge in our entropy-aware trajectory optimization self-distillation framework, enabling reward-model-free policy optimization. Throughout both training stages, OxygenREC-v2 maintains a single unified backbone. We implement OxygenREC-v2 as a 3B-parameter, 1B-activated MoE and deploy it on JD.com's large-scale e-commerce platform. Across multiple online A/B tests, OxygenREC-v2 improves user click-through conversion rate (UCTCVR) by 1.6--4.4% and GMV by 2.8--6.8% over OxygenREC-v1.

Problem

• generative recommendation
• behavior signals
• discriminative objectives
• reward alignment
• user interaction

Innovation

• Generative Recommendation
• Internalized Discrimination
• Behavior Conditioning
• Trajectory Optimization
• Reward-Model-Free RL