PSG: Pair-Space Generation for Efficient Generative Reranking

In modern recommender systems, the Generator-Evaluator (G-E) framework is used for list-wise reranking. However, standard Auto-Regressive (AR) generators suffer from two primary bottlenecks: high latency due to step-by-step generation and significant error accumulation caused by the teacher-forcing training paradigm. The authors investigate whether the generation horizon can be reduced without sacrificing the model's ability to capture complex inter-item dependencies.

The authors propose Pair-Space Generation (PSG), which reformulates the generation process by elevating the "atom" of generation from individual items to ordered item pairs. By treating each pair as a single token, the system generates a sequence of length L/2 instead of L. To handle the resulting quadratic increase in vocabulary size, the authors implement a dynamic pair-token representation module that computes embeddings on-the-fly using a pre-trained encoder. This allows the generator to remain flexible across different candidate sets without needing a massive, static embedding table. The model is trained using a combination of next-token prediction and reinforcement learning (GRPO) to align generated sequences with user engagement metrics.

PSG offers three key advantages:

1. Theoretical Equivalence: The mapping between item-space and pair-space is bijective, meaning PSG maintains the same expressive power as traditional item-space generators.
2. Efficiency: By halving the decoding steps, PSG achieves a theoretical 2x to 4x speedup, with empirical results showing a 1.83x speedup in real-world industrial environments.
3. Robustness: The authors prove a nearly 4x improvement in the worst-case suboptimality bound, as the quadratic dependence of error compounding is mitigated by the shorter decoding path.

In production deployment on Kuaishou, serving over 400 million daily active users, PSG delivered a 0.178% increase in per-user stay time, demonstrating that the efficiency gains translate directly into improved user experience.