研究监管规则、证据与约束条件下的语言模型推理和决策系统。I study language-model reasoning and decision systems under regulatory rules, evidence, and constraints.
我的研究聚焦监管智能、监管规则解析、结构化代码生成与人工智能评测，尤其关注如何从跨行业复杂文档中检索证据、解析规则结构，并将其转化为可执行、可验证和便于人工复核的系统。我曾在汇丰从事监管科技、文档智能与跨境支付智能系统开发。My work spans regulatory intelligence, rule parsing, structured code generation, and AI evaluation. I focus on retrieving evidence from complex cross-domain documents, parsing rule structures, and translating them into executable and verifiable systems that remain open to human review. I previously worked at HSBC on regulatory technology, document intelligence, and intelligent cross-border payment systems.
我关注复杂行业与机构场景中的证据、规则和决策，并尝试把这些问题表述为可检验的人工智能、信息检索、代码生成与评测任务。I study evidence, rules, and decisions in complex industry and institutional settings, formulating them as testable problems in AI, information retrieval, code generation, and evaluation.
如何从长篇法规、业务材料与用户问题中找到相关证据，并识别主体、条件、约束、例外及规则适用范围。How can systems retrieve relevant evidence from long regulations and business materials, then identify subjects, conditions, constraints, exceptions, and rule applicability?
如何将自然语言、监管条文和验证器反馈转化为可执行程序，并利用编译、测试和依赖证据引导生成与修复。How can natural language, regulatory text, and verifier feedback be translated into executable programs, with compiler, test, and dependency evidence guiding generation and repair?
如何区分模型真正利用了证据与推理计划，还是只依赖格式、继承错误脚手架，或产生表面上的分数提升。How can evaluation distinguish genuine use of evidence and reasoning plans from format dependence, corrupted scaffolds, or superficial score gains?
已发表或录用论文、审稿中手稿，以及正在准备的工作。作者列表中我的姓名以粗体标出。Published or accepted papers, manuscripts under review, and work in preparation. My name is shown in bold.
* 共同贡献；† 通讯作者。审稿中与准备中的工作状态依据当前简历列出。* Equal contribution; † corresponding author. Review and preparation status follows the current CV.
以下项目强调我承担的研究与系统工作，而不是单纯罗列产品功能。The projects below emphasize my research and system contributions rather than product features alone.
我领导并参与了一条围绕监管规则理解与合规自动化的连续研究线：从 FinTextQA 的长文档证据检索，到 Span-Grounded Deontic Trees 的规范范围解析，再到 Compliance-to-Code 的法规代码生成、检查与审计报告。核心工作是把主体、条件、约束、例外和证据要求显式表示，并连接到可执行的合规逻辑。I led and contributed to a continuous research line on regulatory rule understanding and compliance automation, spanning long-document evidence retrieval in FinTextQA, normative scope parsing with Span-Grounded Deontic Trees, and executable regulation-to-code generation, checking, and audit reporting in Compliance-to-Code. The central contribution is making subjects, conditions, constraints, exceptions, and evidence requirements explicit and executable.
我从 DeKeyNLU 的任务分解与关键词提取出发，把结构化生成扩展到监管文本，再进一步研究 RepoTrace：将编译、构建、依赖、API 版本和测试失败信息组织为可追踪证据，用于仓库级代码生成与修复。Starting from task decomposition and keyword extraction in DeKeyNLU, I extended structured generation to regulatory text and then to RepoTrace, which organizes compiler, build, dependency, API-version, and test-failure signals as traceable evidence for repository-level code generation and repair.
我共同领导 HARMONICBENCH 的研究设计，使用固定目标和反事实条件比较有帮助、被污染以及依赖格式的推理计划，从而测量模型是否真正利用计划内容，还是仅保留格式或继承错误脚手架。I co-led the research design of HARMONICBENCH, using fixed targets and counterfactual conditions to compare helpful, corrupted, and format-dependent plans, measuring whether models genuinely use plan content or merely retain formatting and inherit flawed scaffolds.
在汇丰全球支付解决方案团队，我负责跨境支付路径规划系统的核心算法与端到端系统设计，将图建模、路径优化、成本与时间预测以及业务约束整合为可部署的多目标决策流程。At HSBC Global Payments Solutions, I led the core algorithmic and end-to-end system design of a cross-border payment-routing system, integrating graph modeling, route optimization, cost and time prediction, and business constraints into a deployable multi-objective decision workflow.
导师：熊辉教授、袁子轩教授；全额奖学金。Supervisors: Prof. Hui Xiong and Prof. Zixuan Yuan; full scholarship.
导师：梁俊卫教授；优秀硕士毕业生；全额奖学金。Supervisor: Prof. Junwei Liang; Outstanding Master’s Graduate; full scholarship.
负责跨境支付路径规划与优化系统。Led Payment AutoPilot for cross-border payment routing and optimization.
研究文档智能、NL2SQL、RAG 与业务预测模型。Worked on document intelligence, NL2SQL, RAG, and predictive models for business analytics.
将媒体与荣誉保持在适当权重，同时补充报告、教学和学术服务。Media and recognition are presented alongside talks, teaching, and academic service.
这篇报道围绕 DeKeyNLU / DeKeySQL 的“先理解、再生成”思路展开，介绍了任务分解、关键词提取、实体检索与 SQL 生成之间的关系。The coverage discusses the “understand first, generate second” approach behind DeKeyNLU / DeKeySQL, including task decomposition, keyword extraction, entity retrieval, and SQL generation.
本摘要依据简历中的报道信息与正式论文整理，并非媒体原文。原始文章直达链接尚未在页面中确认。This summary is compiled from the CV entry and the formal paper; it is not a reproduction of the media article. A verified direct link to the original article is not currently included.