AURA：效用引导的自适应推理实现高效文本检索
原标题：Learning When to Think: Utility-Guided Adaptive Reasoning for Efficient Text Retrieval
AI 摘要
该论文提出 AURA 框架，将面向检索的推理建模为效用引导的计算分配问题，直接监督中间推理状态对相关文档与困难干扰项的区分能力，并训练轻量级效用估计器在推理时自适应决定推理深度。在 BRIGHT 基准上，AURA-4B 以平均 2.8 步推理、每查询 96 个生成 token 达到 34.7 的 nDCG@10，自适应停止将延迟从 4.6 秒降至 0.76 秒；在 MS MARCO 上性能与强推理训练检索器相当，未出现明显域内退化。
正文节选
Learning When to Think: Utility-Guided Adaptive Reasoning for Efficient Text Retrieval Abstract Reasoning-conditioned embeddings can improve retrieval by progressively refining semantic constraints. However, additional reasoning may introduce over-generalized or semantically noisy abstractions that weaken retrieval specificity. Existing methods largely supervise reasoning through final retrieval outcomes, providing limited guidance over the refinement process itself. We propose AURA, a framework