通过轨迹微调让小型语言模型学习检索动作
原标题：The Fellowship of the Query: Learning Retrieval Actions
AI 摘要
该论文研究通过轨迹微调提升小语言模型（SLM）作为检索增强问答中「下一步动作控制器」的能力。作者从教师模型 Qwen 3-Next 80B-A3B Instruct 的搜索轨迹中构建七类动作预测任务，并用 LoRA 监督微调多个 SLM 与 xSLM。在 1,646 个留出动作样本上，Granite 4.1 3B 的 macro-F1 达 0.6536，远高于零样本提示的 0.1736 和 TF-IDF 逻辑回归基线的 0.5399；端到端评估中 Exact Match 从 0.7530 提升至 0.7946，token F1 从 0.7783 提升至 0.8295。
正文节选
The Fellowship of the Query: Learning Retrieval Actions Abstract. Retrieval-augmented question answering requires control decisions about when to decompose a question, search, reformulate, extract evidence, synthesize facts, verify progress, and stop. We study whether trajectory fine-tuning can improve small language models (SLMs) as next-action controllers. We additionally evaluate a low-resource setting in which a single SLM serves as both the controller and the final-answer generator. From ac