Xiang Zhang

关于我About me

我目前在上海交通大学攻读计算机专业硕士,预计于 2028 年 3 月毕业,本科就读于同济大学土木工程专业。我的研究兴趣是 LLM Agents、Agentic Memory 和 Data-Centric AI。I am a master’s student in Computer Science at Shanghai Jiao Tong University, expecting to graduate in March 2028. I received my B.Eng. in Civil Engineering from Tongji University. My research interests are LLM Agents, Agentic Memory, and Data-Centric AI.

研究与论文Research & Publications

VLDB 2026VLDB 2026Dial: A Knowledge-Grounded Dialect-Specific NL2SQL System — overview

Dial: A Knowledge-Grounded Dialect-Specific NL2SQL System

Xiang Zhang, Hongming Xu, Le Zhou, Wei Zhou, Xuanhe Zhou, Guoliang Li, Yuyu Luo, Changdong Liu, Guorun Chen, Jiang Liao, Fan Wu

PVLDB 2026第一作者First author2025.07 — 2026.03

Dial 面向不同数据库的 SQL 方言,将规划、检索、生成、执行反馈和语义校验串联起来,无需模型微调即可生成跨环境代码,减少知识幻觉、方言干扰和不可执行的输出。NL-LQP 把自然语言需求分解为带依赖关系的 DAG Macro-Operators;分层知识库 HINT-KB 从官方文档整理函数知识与隐式执行约束,为各个步骤提供面向意图的 Structured RAG。

生成后,系统根据真实执行反馈检索约束并定向修复,用 Logical Plan 检查语义偏移(Semantic Drift),再将验证后的经验写回知识库。我们构建了覆盖 6 个数据库、2,218 个样例的 DS-NL2SQL Benchmark;相比最强标准 Baseline,执行准确率平均提升 10.25%,方言特征覆盖率平均提升 15.77%。

Dial combines planning, retrieval, generation, execution feedback, and semantic validation to generate SQL across database dialects without model fine-tuning, addressing knowledge hallucinations, dialect interference, and non-executable outputs. NL-LQP decomposes a natural-language request into DAG macro-operators with dependencies. The hierarchical HINT-KB organizes function knowledge and implicit execution constraints from official documentation to support intent-oriented structured RAG.

The system retrieves constraints and makes targeted repairs using actual execution feedback, checks logical plans for semantic drift, and writes validated experience back to the knowledge base. We built DS-NL2SQL with 2,218 test cases across 6 databases. Compared with the strongest standard baseline, Dial improves average execution accuracy by 10.25% and dialect feature coverage by 15.77%.

arXivarXivMemTrace: State-Consistent Memory for Long-Horizon Coding Agents — overview

MemTrace: State-Consistent Memory for Long-Horizon Coding Agents

Hongming Xu, Le Zhou, ZhongHe Jin, Xiang Zhang, Bo Tang, Zhiyu Li, Xuanhe Zhou, Juncheng Zhang

arXiv2026.10

MemTrace 研究长程 Coding Agent 在上下文刷新后如何恢复记忆,同时保持记忆与当前代码状态一致。它用不可变的 Memory Traces 和 Memory Trace Graph 保留执行历史,再根据当前仓库状态校验历史证据,避免恢复已经失效的信息。

在相同 backbone 和 harness 下,论文报告 Codex CLI 在三个长程编码基准上的提升:DeepSWE pass@1 +21.2 个百分点,SWE-EVO Resolved Rate +4.4 个百分点,SWE-Milestone Score +17.8 个百分点。

MemTrace studies how long-horizon coding agents can recover memory after a context refresh while keeping it consistent with the current code state. Immutable memory traces and a memory trace graph preserve execution history; validation against the current repository state checks that recovered evidence is still valid.

With the same backbone and harness, the paper reports gains under Codex CLI on three long-horizon coding benchmarks: +21.2 percentage points in DeepSWE pass@1, +4.4 points in SWE-EVO Resolved Rate, and +17.8 points in SWE-Milestone Score.

项目经历Projects

Agentic MemoryAgentic Memory

2026.04 — 2026.08

面向长程 LLM Agent 的记忆系统调研与评测Survey and evaluation of memory systems for long-horizon LLM agents

针对长程 LLM Agent 的上下文膨胀、历史检索和知识冲突,从数据管理角度梳理记忆系统的 存储、提取、检索与维护,比较向量、结构化、知识图谱和混合记忆架构。

建立统一比较框架,从检索、动态更新、长程稳定性和成本四个维度分析主流系统,整理增量更新、冲突消解、选择性遗忘和 Memory-to-Action 等开放问题。

Surveyed memory systems for long-horizon LLM agents facing context growth, history retrieval, and knowledge conflicts. From a data-management perspective, organized the field around storage, extraction, retrieval, and maintenance, comparing vector, structured, knowledge-graph, and hybrid architectures.

Established a shared comparison framework covering retrieval, dynamic updates, long-horizon stability, and cost; identified open questions in incremental updates, conflict resolution, selective forgetting, and memory-to-action.

项目演示Demo

AI 技术日报AI Daily Digest

社媒观点 · 技术实践 · 长期观察Community perspectives · Technical practice · Ongoing analysis

把更多篇幅留给 X、Reddit 和技术博客里的评价、争论与实践经验。同一事件保留不同人的说法,回到原始材料核对;产品发布和热度消息只作简讯。An AI digest focused on opinions, debates, and hands-on experience from X, Reddit, and technical blogs. Different takes on the same event are preserved and checked against original sources; launches and trending news stay brief.

编辑观察从知识库的旧线索出发,联系新材料提出判断,再把依据、反例和修正留下来。让每天的阅读接得上过去,也让自己的看法经得起回看。Editorial analysis connects earlier knowledge-base entries with new evidence. Arguments, counterexamples, and revisions are kept for later review, so each issue builds on previous reading and earlier judgments remain open to correction.

个人经历Experience

上海交通大学Shanghai Jiao Tong University

2025.09 — 2028.03

计算机专业 · 硕士研究生(在读)Computer Science · Master's student

导师Advisor: 周煊赫Xuanhe ZhouTheseus Labs

预计 2028 年 3 月毕业Expected graduation: March 2028

同济大学Tongji University

2021.09 — 2025.06

土木工程 · 工学学士Civil Engineering · B.Eng.

同济大学优秀学生 · 同济大学优秀毕业生Outstanding Student · Outstanding Graduate, Tongji University