EVENT DOSSIER
The research team proposed a single-step decoding strategy called SPD (Single Forward Pass), aimed at addressing the issue of slow inference speed in generative reordering tasks. This strategy utilizes a lightweight self-attention head to directly extract the score matrix from the pre-filled hidden states of large language models (LLMs), and employs the Hungarian algorithm to directly calculate the optimal binary matching, thereby outputting N ordinal numbers in a single forward pass with an O(1) time complexity. Experiments show that this method, combined with LoRA fine-tuning and autoregressive distillation training, improves the inference speed from traditional methods to 28 milliseconds (a 64-fold increase), while maintaining the same ranking quality as the teacher model. Complete ablation experiments further verified the specific contributions of architecture, training signals, and backbone adaptation to performance. This framework successfully combines generative ranking with combinatorial optimization, providing a new technical approach for constructing other O(1) decoding mechanisms to support real-time ranking.
提出 SPD（Single Forward Pass）解码策略，将生成式重排序的推理速度提升 64 倍至 28 毫秒。该策略利用轻量级自注意力头从 LLM 预填充隐藏状态中读取 $N \times K$ 得分矩阵，并通过匈牙利算法直接计算最优二分匹配以输出 $N$ 个序数，实现 $O(1)$ 前向传播。研究通过 LoRA 微调结合自回归蒸馏训练，在保持与教师模型排名质量相当的同时达成该性能。完整消融实验分解了架构、训练信号及骨干适配的贡献，该框架将生成式排名与组合优化连接，为实时排名提供其他 $O(1)$ 解码机制路径。