信息增益校准的深度搜索智能体在线策略蒸馏
原标题：Information-Gain Calibrated On-Policy Distillation for Deep Search Agents
AI 摘要
该论文针对深度搜索智能体的在线策略蒸馏（OPD）提出改进方法 IG-OPD。作者指出标准 OPD 存在“证据盲监督”（EBS）问题：监督仅基于检索前上下文，导致预算先验按 token 量分配监督、方向先验偏向教师偏好，而忽略检索到的证据实际效用。IG-OPD 利用检索后信息增益校准这两个先验，在六个多模态搜索基准上使 2B 和 4B 学生模型平均准确率分别提升 3.4 和 1.7 个百分点，4B 学生在 MMSearch 和 BrowseComp-VL 上超过其 8B 教师模型。
正文节选
Information-Gain Calibrated On-Policy Distillation for Deep Search Agents Abstract On-policy distillation (OPD) has recently emerged as a promising approach for transferring the capabilities of large agents to compact students. In deep search, where agents iteratively retrieve external evidence, a teacher’s preference among plausible queries before retrieval may not predict which queries will return useful evidence. Standard OPD nevertheless supervises each student query from the preceding conte