DecisionRank：面向决策模型重排器的同策略蒸馏
原标题：Paper page - DecisionRank: On-Policy Distillation for Decision-Model Rerankers
AI 摘要
该论文提出 DecisionRank，一种通过 item-space on-policy distillation（OPD）训练决策模型重排器的方法：将学生模型对文档的打分排序成序列，由生成式教师模型基于该排序列表给出条件反馈，直接监督这些分数，并结合离线排序监督。在 TREC DL、BEIR、BRIGHT 等检索基准上取得 SOTA，1.7B 参数版本在 TREC DL19-23 上超过其 27B 生成式教师，8B 版本推理速度比生成式 listwise 方案快 24.5 倍（TREC DL）和 16.7 倍（BRIGHT）。
正文节选
DecisionRank: On-Policy Distillation for Decision-Model Rerankers Abstract DecisionRank uses item-space on-policy distillation to transfer ranking knowledge from generative teachers to efficient decision models, achieving state-of-the-art retrieval performance with significantly faster inference. Decision models target structured tasks by directly predicting choices or scores, avoiding autoregressive decoding latency. However, transferring a generative LLM's ranking knowledge to a decision model