动态模型路由Adaptive LLM Routing
结合任务上下文、Agent 角色与执行反馈,动态决定何时调用不同能力的模型,在任务成功率与计算开销间取得更优平衡。Using context, agent roles, and execution feedback to select models at each step, balancing task success against computation costs.
让多智能体系统更高效、更可靠。Towards more efficient and reliable multi-agent intelligence.
中国石油大学(华东)计算机技术硕士研究生。聚焦大语言模型多智能体系统、角色自适应模型路由、智能体记忆与协作优化,并探索 AI 在工业场景中的应用。Master's student in Computer Technology at China University of Petroleum (East China). My work focuses on LLM-based multi-agent systems, role-adaptive routing, agent memory, and practical AI for industrial systems.
从算法研究到工程系统,关注能够落地的智能体技术。Bridging method development and real-world AI systems.
我目前就读于中国石油大学(华东),攻读计算机技术硕士学位,本科毕业于中北大学计算机科学与技术专业。我的研究围绕 LLM-based Multi-Agent Systems 展开,重点探索动态模型路由、角色感知决策、执行风险评估与智能体协作机制。
I am pursuing a master's degree in Computer Technology at China University of Petroleum (East China), after earning a bachelor's degree in Computer Science and Technology from North University of China. My research explores LLM-based multi-agent systems, with an emphasis on adaptive model routing, role-aware decision-making, execution-risk estimation, and collaborative intelligence.
除了科研工作,我也参与工业智能化项目,尝试将知识图谱、RAG、多模态感知与 Agent 工作流结合,服务于复杂装备的监测、维护和智能化分析。
Beyond foundational research, I investigate applied AI for industrial workflows, integrating knowledge graphs, RAG, multimodal sensing, and agent-based automation for monitoring and maintenance.
从模型选择、协作记忆到实际系统的统一研究路径。From model selection and collaborative memory to intelligent applications.
结合任务上下文、Agent 角色与执行反馈,动态决定何时调用不同能力的模型,在任务成功率与计算开销间取得更优平衡。Using context, agent roles, and execution feedback to select models at each step, balancing task success against computation costs.
研究不同角色间的经验共享、记忆使用与协作决策机制,探索复杂工作流中可复用的团队知识。Studying experience sharing, memory usage, and collaboration across specialized agent roles in complex workflows.
将知识增强生成、多模态数据融合与 Agent 工作流引入设备监测、故障诊断和专业分析任务。Combining grounded generation, multimodal fusion, and agent workflows for equipment monitoring, fault analysis, and domain tasks.
面向异构大模型的多智能体动态路由研究。Efficient routing for heterogeneous language models in multi-agent settings.
提出轻量级角色自适应模型路由框架:利用在线 Critic 评估和故障归因构建六维执行风险状态,再由角色条件超网络预测不同模型的执行风险;在线 MemoryBank 辅助决策。面向 SWE-bench Lite 与 WebArena-Verified 评估。
A lightweight role-adaptive router that derives six-dimensional execution-risk states from online Critic evaluations and forensic attribution. A role-conditioned hypernetwork estimates model-specific risks, supported by an online MemoryBank. Evaluated on SWE-bench Lite and WebArena-Verified.
提出面向图结构多智能体系统的轻量级动态路由方法,通过语义嵌入与结构元特征组成双信号路由器,并采用从冷启动到迭代进化的优化过程,利用路由失败反馈更新决策。
A context-aware dual-signal router for graph-based MAS, combining semantic embeddings with structural meta-features. A cold-start-to-iterative-evolution training strategy learns from routing failures using on-policy negative feedback.
科研系统、行业智能化与软件工程实践并重。Research systems, domain-specific AI, and software engineering practice.
围绕修井作业设计文档自动生成与智能运维,将井下小修领域知识库、知识图谱、LLM、RAG 与多智能体工作流结合;融合热成像、三轴振动、PLC 及四气体监测数据,研究异常设备识别、故障预警与设备健康报告生成。
Research on automated job-plan generation and intelligent maintenance for well servicing equipment. Combines domain knowledge, knowledge graphs, LLM/RAG, and agent workflows with thermal imagery, 3-axis vibration, PLC, and four-gas sensing for fault detection and equipment-health reporting.
研究方案拟融合领域微调 SAM、知识图谱与 LLM,构建目标驱动的自主规划、工具优选和多智能体协作机制,辅助岩石孔隙结构表征、渗流模拟与物性分析。
A proposed agentic framework combining domain-adapted SAM, knowledge graphs, and LLMs for goal-driven workflow planning, tool selection, and multi-agent collaboration in pore characterization, flow simulation, and petrophysical analysis.
基于 LangGraph、LangChain 和 Qwen 开发多智能体数据研究系统。通过假设生成、流程控制、检索、代码执行、可视化、报告生成和质量审查七类角色协作,实现端到端数据分析。
An agent-based data research platform built with LangGraph, LangChain, and Qwen. Seven specialized roles—Hypothesis, Process, Search, Code, Visualization, Report, and QualityReview—collaborate on end-to-end analysis.
以中北大学计算机科学与技术学院为场景,通过 LangChain 和百度千帆平台建立学院规章制度及文化资料向量知识库,提供可溯源的领域知识问答。
An undergraduate project using LangChain and Baidu Qianfan to ground LLM responses in a vector database of college policies and institutional materials, enabling domain-specific question answering.
独立完成毕业实习项目的数据库、前后端和 Apriori 算法开发,基于现有数据集挖掘商品关联规则,实现推荐功能和完整的数据增删改查。
Independently implemented database, frontend, backend, and Apriori association-rule algorithms for product recommendation, with full CRUD functionality.
面向个人与团队的日程管理与协作工具,集成会议纪要、快捷笔记、日程安排及智能推荐模块,改善日常任务组织体验。
A productivity and collaboration platform featuring meeting minutes, quick notes, scheduling, and intelligent recommendation modules for individuals and teams.
在自主学习 Windows 逆向工程的过程中实现的工具,可解析 Portable Executable 文件结构、节表、导入导出表、重定位表与符号地址等信息。
A systems-learning project for inspecting Portable Executable structures, including sections, imports, exports, relocations, and symbol addresses.
以计算机科学为基础,持续探索可信、高效的智能体系统。A computer science background with a focus on intelligent agent systems.
新加坡 · 可信智能体相关研究(拟进行;具体时间与资格以正式审批为准)Singapore · Proposed visiting research on trustworthy agents; subject to formal approval
计算机技术 · 硕士研究生 · GPA 4.08 / 5(中文简历)Master's student in Computer Technology · GPA 4.08 / 5 (Chinese résumé)
计算机科学与技术 · 本科 · 本科 GPA 3.62(英文 CV)B.S. in Computer Science and Technology · Undergraduate GPA 3.62 (English CV)
竞赛、科研实践及在校表现方面的主要荣誉。Selected competition awards and university recognitions.
欢迎围绕 LLM、多智能体系统、Agent 路由与工业 AI 展开学术交流或项目合作。I welcome research discussions and collaborations around LLMs, multi-agent systems, model routing, and applied industrial AI.