马义然

+86 177 1399 9172·yiranrexma@outlook.com·yiranrexma.monticule.tech

基础语言模型 · 持续学习 | 面壁垂域基座经验

证件照

教育背景

北京大学硕士研究生·外国语言学及应用语言学 2026.09 –
北京邮电大学本科·英语(语言学与语言智能)·辅修·数据科学与大数据技术 2022.09 – 2026.06
  • 专业排名 1/44·GPA 3.87/4.0·优秀本科毕业论文(前 2%)

实习与科研经历

清华大学计算机系·自然语言处理实验室 × 面壁智能(THUNLP × ModelBest)|科研实习生 2024.10 – 2025.09
  • 深度参与最高人民法院“法信”法律基座大模型的司法仿真环境合规评测建设
  • 搭建多智能体博弈庭审框架 LinguaGame:对话质量在清晰度、简洁性、论证性、策略性四个维度提升 6.3%–20.1%
  • 提出法律文风合规评测方法 CLASE:与人工评审一致性较传统评估方法提升 94.6%–163.0%
北京邮电大学·叶培大创新创业学院|学生科研项目负责人 2023.10 – 2025.02
  • 主持构建中英图文双关翻译框架 Pun2Pun:系统评测 4 个模型家族的 9 款模型;CVO 翻译策略使翻译质量较模型标准输出提升 15.7%–205.1%
  • 主持零样本多智能体混合表格推理框架 PanelTR:准确率较最强前沿模型高 9.2 个百分点、超越监督式方法;延伸实证研究发现多智能体反思的"过度思考阈值"

核心技能

  • 大模型:PyTorch / Transformers·领域微调(SFT)与强化学习·持续学习·多模态训练·Agent / Harness 工程·评测
  • 语言:中文(母语)·英语(专业八级 89 优秀、IELTS 7.5)

荣誉奖项

  • MCM/ICM Meritorious Winner(前 7%,2025)·大学生创新创业训练计划国家级项目·英国剑桥大学访学优秀学生代表

论文成果

Yiran Rex Ma*, Yuxiao Ye*, and Huiyuan Xie. CLASE: A Hybrid Method for Chinese Legalese Stylistic Evaluation. LREC 2026.

Yuxiao Ye, Yiming Zhang*, Yiran Rex Ma*, Huiyuan Xie, Huining Zhu, and Zhiyuan Liu. LinguaGame: A Linguistically Grounded Game-Theoretic Paradigm for Multi-Agent Dialogue Generation. ACL 2026 Findings.

Yiran Rex Ma, Shan Huang, Yuting Xu, Ziyu Zhou, and Yuanxi Wei. Pun2Pun: Benchmarking LLMs on Textual-Visual Chinese-English Pun Translation via Pragmatics Model and Linguistic Reasoning. ACL 2025 SRW.

Yiran Rex Ma. Do Androids Question Electric Sheep? A Multi-Agent Cognitive Simulation of Philosophical Reflection on Hybrid Table Reasoning. ACL 2025 SRW.

Yiran Rex Ma. PanelTR: Zero-Shot Table Reasoning Framework Through Multi-Agent Scientific Discussion. IJCNN 2025.

李霜洁, 马义然, 孙茂松, 刘知远. 基于图结构和深度学习的古文字识别/缺损字形重建方法和装置: CN202510989823.X/9824.4.

* 同等贡献

Yiran Rex Ma

+86 177 1399 9172·yiranrexma@outlook.com·yiranrexma.monticule.tech

Foundation language models & continual learning | Domain foundation model experience with ModelBest

Education

Peking University — M.A. in Foreign Linguistics and Applied Linguistics Sep 2026 –
Beijing University of Posts and Telecommunications — B.A. in English, Linguistics and Language Technologies; Minor in Data Science Sep 2022 – Jun 2026
  • GPA 3.87/4.0, ranked 1/44; Outstanding Undergraduate Thesis (top 2%)

Experience

Tsinghua University, THUNLP × ModelBest — Research Intern Oct 2024 – Sep 2025
  • Deeply involved in building the judicial simulation environments and compliance evaluation of Faxin, the Supreme People's Court legal foundation model
  • Built LinguaGame, a multi-agent game-theoretic courtroom framework: dialogue quality improved by 6.3%–20.1% across clarity, conciseness, argumentation, and tactic
  • Proposed CLASE, a legalese compliance-evaluation method: alignment with human experts improved by 94.6%–163.0% over conventional evaluation methods
BUPT, Ye Peida College of Innovation and Entrepreneurship — Student Project Lead Oct 2023 – Feb 2025
  • Built Pun2Pun, a Chinese-English textual-visual pun translation framework benchmarking 9 models from 4 families; the CVO strategy lifts translation quality by 15.7%–205.1% over standard model output
  • Proposed PanelTR, a zero-shot multi-agent hybrid-table reasoning framework: +9.2 points over the strongest frontier model, surpassing supervised methods; follow-up work identified an "overthinking threshold" in multi-agent reflection

Publications

Yiran Rex Ma*, Yuxiao Ye*, and Huiyuan Xie. CLASE: A Hybrid Method for Chinese Legalese Stylistic Evaluation. LREC 2026.

Yuxiao Ye, Yiming Zhang*, Yiran Rex Ma*, Huiyuan Xie, Huining Zhu, and Zhiyuan Liu. LinguaGame: A Linguistically Grounded Game-Theoretic Paradigm for Multi-Agent Dialogue Generation. ACL 2026 Findings.

Yiran Rex Ma, Shan Huang, Yuting Xu, Ziyu Zhou, and Yuanxi Wei. Pun2Pun: Benchmarking LLMs on Textual-Visual Chinese-English Pun Translation via Pragmatics Model and Linguistic Reasoning. ACL 2025 SRW.

Yiran Rex Ma. Do Androids Question Electric Sheep? A Multi-Agent Cognitive Simulation of Philosophical Reflection on Hybrid Table Reasoning. ACL 2025 SRW.

Yiran Rex Ma. PanelTR: Zero-Shot Table Reasoning Framework Through Multi-Agent Scientific Discussion. IJCNN 2025.

Shuangjie Li, Yiran Rex Ma, Maosong Sun, and Zhiyuan Liu. Method and Apparatus for Recognizing Ancient Scripts and Reconstructing Damaged Glyphs Based on Graph Structure and Deep Learning. CN202510989823.X/9824.4 (Invention Patent).

* Equal contribution

Skills

  • LLM engineering: PyTorch / Transformers, SFT & RL, continual learning, multimodal training, agent & harness engineering, evaluation
  • Languages: Mandarin Chinese (native); English (TEM-8 Excellent 89, IELTS Academic 7.5)

Honors & Awards

  • Meritorious Winner, MCM/ICM 2025 (top 7%)·National Innovation and Entrepreneurship Training Program·Visiting Student Representative, University of Cambridge