Long-Horizon Tasks Retrieval-Augmented Generation Industrial AI Memory Compression
摘要

针对长周期钢铁设备巡检需推理异构历史记录的挑战,现有检索增强生成系统视日志为静态语料,缺乏诊断价值评估。本文提出 ConMem,一种面向 LLM 辅助巡检的贡献感知记忆框架。该方法将日志分割为功能证据单元,利用 Shapley 风格估算各单元对下游诊断的贡献,并在受限预算下保留高价值证据。实验表明,ConMem 在真实数据集上 QA 准确率达 76.0%,相比基线减少 88.2% 输入令牌及 86.6% 响应时间,能有效保留微弱早期风险信号。

AI 推荐理由

论文提出贡献感知记忆框架,核心解决长周期任务中的记忆选择与价值评估问题。

研究机构
Shandong University, School of Software, Jinan, Shandong, China Boston University, Metropolitan College, Boston, MA, USA Nanjing University of Science and Technology, School of Computer Science and Engineering, Beijing, Peking, China National University of Singapore, Department of Statistics and Data Science, Singapore, Singapore Joint SDU-NTU Centre for Artificial Intelligence Research (C-FAIR), Shandong University, Jinan, Shandong, China Ruihao Steel Holding Group Co., Ltd
论文信息
作者 Bingchen Liu, Yuanyuan Fang, Lei Liu, Guangyuan Dong, Xing Fu et al.
发布日期 2026-07-30
arXiv ID 2607.28126
相关性评分 9/10 (高度相关)