ENTERPRISE AI ANALYSIS

Methodology Details: Gen-Reranker & Training

GReF's methodology comprises a three-part architecture and training strategy.

1. Gen-Reranker Architecture: It consists of a bidirectional transformer encoder that extracts context-aware embeddings for candidate items, and a dynamic autoregressive decoder. The decoder generates next-item representations and dynamically matches them with candidate embeddings, eliminating the need for a full vocabulary output layer and handling billions-scale item pools.

2. Pre-training on Recommender World Knowledge: Gen-Reranker is initially pre-trained on vast unlabeled item exposure order data from the recommendation system. This cross-entropy based pre-training (Equation 6) efficiently initializes model parameters, captures broad user interests, and enhances generalization by learning from high-quality sequences generated by existing recommendation systems.

3. Post-training on User Preferences with Rerank-DPO: To integrate explicit user preferences and enable end-to-end optimization, GReF uses Rerank-DPO. Preference pairs are constructed based on item personalization scores (Equation 7), which consider original exposure position and user feedback (e.g., clicks). The DPO objective (Equation 8) then directly aligns the model with preferred sequences without an explicit reward model.

4. Ordered Multi-token Prediction (OMTP): For efficient inference, OMTP trains Gen-Reranker with n output heads to predict multiple future items simultaneously (Equation 9). An additional pairwise loss (Equation 10) is applied to ensure these multi-token predictions maintain their correct sequential order, crucial for causal modeling of user behavior. This significantly reduces forward passes and latency during real-time inference.

Experimental Results: Superior Performance & Efficiency

Extensive offline experiments on both public (Avito) and industrial (Kuaishou) datasets demonstrated GReF's superior performance. On Avito, GReF achieved the highest AUC (0.7384, ~1.5% higher than best baseline) and NDCG (0.7478, ~0.8% higher than PIER). On Kuaishou, it recorded AUC of 0.7387 (~1.4% higher) and NDCG of 0.7498 (~0.5% higher), consistently outperforming state-of-the-art reranking methods. Crucially, GReF achieved an inference time of 12.97ms on Kuaishou, nearly comparable to non-autoregressive models (NAR4Rec at 12.67ms) and over 4x faster than Seq2Slate, validating OMTP's efficiency.

Online A/B tests on Kuaishou's video app (over 300 million DAU) further confirmed GReF's effectiveness, showing significant improvements: +0.33% in Views, +0.42% in Long Views, +1.19% in Likes, +2.98% in Forwards (shares), and +1.78% in Comments. These results highlight GReF's robustness, accuracy, and practical deployability in real-world, high-scale recommendation systems.

Discussion: Bridging the Gap

The findings from GReF's offline and online evaluations underscore its ability to overcome long-standing challenges in reranking. The strong performance metrics across diverse datasets (Avito and Kuaishou) attest to its robust generalization capabilities, likely attributed to the comprehensive pre-training on real-world exposure data. The ablation studies confirm the synergistic effect of pre-training for parameter initialization and Rerank-DPO for preference alignment, demonstrating that neither stage alone achieves optimal results and that direct DPO application can lead to instability without pre-training. The most significant practical implication is the success of OMTP in bridging the gap between generative model expressiveness and real-time efficiency. By enabling parallel, order-preserving multi-token prediction, GReF achieves inference speeds competitive with non-autoregressive models while retaining the causal modeling benefits of autoregressive generation. This makes GReF a viable solution for industrial-scale recommendation systems where both high accuracy and low latency are paramount, previously a trade-off that constrained generative reranking models.