沈瑚双 Hushuang Shen

2027 年保研去向：中国科学院大学 · 计算技术研究所
计算机科学与技术 · UCAS / ICT · 尚未入学

我是中国石油大学（北京）数据科学与大数据技术专业本科生，预计 2027 年毕业，专业成绩排名 3/94（2026.09）。

研究兴趣包括多模态大模型、Agent 记忆与检索增强。已发表多篇 SCI 论文。

GitHub CV Personal page Research Engineering Notes
Darssal的动漫头像

News

  1. 国家级大学生创新训练项目“基于联合时空注意力机制和多模态大模型的地层信息处理系统”结题，担任项目成员。

  2. Persistent Retrieval in Hierarchical Memory Systems for Long-Context Agents: A Controlled Evaluation 获 IEEE SMC 2026 录用。

  3. Two-Stage Recruitment Text Labeling via Lexicon-Guided Routing and Retrieval-Augmented Large Language Models 以第一作者发表于 JoVE。

  4. FracLogGPT: A Multimodal Large Language Model for Fracture Interpretation in Imaging Logging 以第一作者在线发表于 Electronics，正式引用卷年为 2026。

Research

* 表示共同贡献

FracLogGPT: A Multimodal Large Language Model for Fracture Interpretation in Imaging Logging

Hushuang Shen, Ang Li, Liyan Zhang, Xiangxiang Liu

Electronics 15(1), 127, 2026 · 第一作者 · 已发表

面向成像测井裂缝解释，基于 Qwen2.5-VL-7B 组织 DAPT、LoRA SFT 与 DPO 三阶段适配，联合评价裂缝计数、空间一致性与文本描述。负责方法设计、软件实现、实验分析与初稿撰写。

论文报告：独立测试集上的裂缝计数分类 Count-F1 为 0.70。

Paper ↗ DOI ↗ Code · Lite ↗
BibTeX
@article{shen2026fracloggpt,
  title = {FracLogGPT: A Multimodal Large Language Model for Fracture Interpretation in Imaging Logging},
  author = {Shen, Hushuang and Li, Ang and Zhang, Liyan and Liu, Xiangxiang},
  journal = {Electronics},
  year = {2026},
  volume = {15},
  number = {1},
  pages = {127},
  doi = {10.3390/electronics15010127}
}
Architecture schematic: borehole-image preparation, training-only data split, DAPT, LoRA SFT and expert-ranked DPO preferences, followed by multimodal inference through the inherited Qwen ViT encoder, vision-language merger and LLM decoder.
Data, adaptation & inference ↗

Persistent Retrieval in Hierarchical Memory Systems for Long-Context Agents: A Controlled Evaluation

Hushuang Shen*, Yuan Wang*

IEEE SMC 2026 · 共同一作 · 已录用

固定模型、解码、提示骨架、检索评分与预算策略，在 LoCoMo、LongBench v2 和 LongMemEval 上比较五种记忆配置，检验持久检索在证据充足与稀缺条件下的作用。

论文报告：证据稀缺协议下，L0+L2 相对当前轮组织基线的平均绝对得分增益为 0.103。

Program ↗ Code · MemLab ↗
BibTeX
@unpublished{shen2026persistent,
  title = {Persistent Retrieval in Hierarchical Memory Systems for Long-Context Agents: A Controlled Evaluation},
  author = {Shen, Hushuang and Wang, Yuan},
  year = {2026},
  note = {Accepted for presentation at IEEE SMC 2026; equal contribution},
  url = {https://conf.papercept.net/conferences/conferences/SMC26/program/SMC26_ContentListWeb_3.html#tua16_05}
}
Architecture schematic: task-scoped L0-L3 memories, optional candidate retrieval, filtering and fixed-score reranking, budgeted prompt assembly and frozen decoding. L3 expansion requires a filtered reference hit; protocol insets compare evidence placement and a strip lists five configurations.
Memory retrieval & execution ↗

Two-Stage Recruitment Text Labeling via Lexicon-Guided Routing and Retrieval-Augmented Large Language Models

Hushuang Shen, Yuan Wang

Journal of Visualized Experiments (231), e70153, 2026 · 第一作者 · 已发表

以 AI 与环保词表分流：仅命中一个领域时直接标为该类，未命中归为 Other；仅将双域命中样本交给检索增强模型，完成三分类标注。

论文报告：在 3,000 条人工核验样本上准确率为 98.47%，并用于约 693 万条招聘信息的语料标注。

Paper ↗ DOI ↗ Baseline ↗
BibTeX
@article{shen2026recruitment,
  title = {Two-Stage Recruitment Text Labeling via Lexicon-Guided Routing and Retrieval-Augmented Large Language Models},
  author = {Shen, Hushuang and Wang, Yuan},
  journal = {Journal of Visualized Experiments},
  year = {2026},
  number = {231},
  pages = {e70153},
  doi = {10.3791/70153}
}
Architecture schematic: offline document chunking and MiniLM embeddings build a LanceDB index. Lexicons route single and no matches to direct labels; only dual matches use query embedding, cosine retrieval, filtered evidence and a fixed prompt with Qwen for classification, with separate retrieval logs.
Knowledge, routing & RAG ↗

以上数值为论文报告的结果。配图为依据论文方法与评估协议重绘的示意图，点击可放大。公开工程代码的实现与验证范围见 Verification & Sources ↗。

Research Experience

基于时序Transformer的跨井域泛化测井预测系统

2026.06—2027.06计划周期

国家级大学生创新训练项目 · 新立项／在研

研究方向：时序 Transformer、跨井域泛化与测井预测。

基于联合时空注意力机制和多模态大模型的地层信息处理系统

2026.09 结题

国家级大学生创新训练项目 · 项目成员

参与 Web 平台前后端开发、地质数据读取与可视化，并围绕裂缝识别开展多模态模型任务验证。

Selected Honors

完整获奖记录见 CV ↗