Curriculum Vitae

Education

• M.S. in Computer Science, National University of Singapore (NUS), 2026.8 – 2027.7 (expected)
• B.S. in Business Analytics, Shanghai University of Finance and Economics (SUFE), 2022.9 – 2026.6
  • GPA: 3.73 / 4.0 (Rank 8 / 128)
  • TOEFL: 101 · GRE: 327 · JLPT: N2

Research & Industry Experience

Tencent Youtu AI Lab — Research Intern (2025.10 – 2026.3)

• Adversarial Data Synthesis. Built an LLM-driven adversarial sample synthesis system with a 5-dimensional evolution strategy (style transfer, scene injection, metaphor substitution, complex logic, multilingual mixing). Generated 100K+ highly stealthy violation samples from seed data with Asyncio-based concurrency and SQLite logging.
• Medusa Unified Moderation Model. Designed a “linear-head + Medusa TextCNN multi-head” architecture on XLM-RoBERTa for 26-dimensional atomic detection, integrating 100+ rules. Achieved 92.3% accuracy on financial violations, −35% false-positives, +60% moderation efficiency, and hundreds of thousands of QPS.
• WeMM-Embedding. Co-developed a multimodal embedding model on a Qwen3-VL backbone with a Deep Fusion module. Reached open-source SOTA on MMEB-V2 and UVRB leaderboards.

SUFE FinAI Center — Research Assistant (2025.6 – present)

• ICBC Scientist-Discovery Agent. Co-developed a RAG + Agent system that mines scientist information, analyzes affiliated enterprises, and predicts sales opportunities. Contributed to FinCorpus (a 100B-token financial corpus).
• Fin-R1 Financial LLM. Co-developed China’s first open-source financial reasoning LLM (7B, based on Qwen2.5-7B-Instruct). Built 60K financial CoT data, trained with SFT + GRPO. Achieved FinEval score 75.2 (2nd place overall).
• Financial Evaluation Suite. Co-designed FinEval (26K problems), Alibaba Cloud financial multimodal evaluation data, and Ant Group’s financial hard-problem evaluation set.

ICBC Head Office, Private Banking Department — Research Intern (2026.1 – present)

• Scientist-Discovery Agent. Built multi-dimensional profiles of high-net-worth scientists from open data (patents, papers, registries). Used Agent + RAG to automate relationship construction, opportunity insight, and personalized “one-customer-one-strategy” marketing.
• Wealth-Advisor Empowerment. Developed a Private-Banking sparring assistant that distills financial knowledge and best practices, simulates real-world scenarios, and produces sales scripts and recommendations.

Publications

• FinVault: Benchmarking Financial Agent Safety in Execution-Grounded Environments. First Co-Author. Under review at ACL 2026.

</ul>

Selected Projects

Fin-R1 — Financial Reasoning LLM with Reinforcement Learning

• Background: Develop a 7B lightweight model for complex financial reasoning across banking, securities, insurance and trust businesses.
• Approach: Distilled 60K financial CoT data from DeepSeek-R1; two-round quality filtering (rules + model scoring + logical consistency); two-stage training (SFT + GRPO); Verifier-augmented reward.
• Results: FinQA / ConvFinQA near SOTA, FinEval 75.2 (2nd place); FinQA 76.0, ConvFinQA 85.0; vLLM one-click deployment.

ICBC — Scientist–Entrepreneur Integrated Agent

• Background: Build an intelligent customer-identification and relationship-construction system for the scientist–entrepreneur community to support tech-transfer business.
• Approach: LLM-based parsing of patents, papers and news; unified scientific knowledge graph and enterprise data; RAG-based expert KB; family-business opportunity insight, dialogue sparring, and one-customer-one-strategy marketing.
• Results: Multi-criteria scientist discovery, panoramic profiles (person/firm/family/society), affiliated-enterprise analysis, automatic marketing-graph generation.

Tencent — Medusa Unified Content-Safety Moderation Model

• Background: Internet content evolves quickly; rule-based systems struggle with new violation patterns.
• Approach: “Linear-head + Medusa TextCNN multi-head” hybrid on XLM-RoBERTa; partial-label masking for sparse labels; 10 prompt templates for few-shot learning.
• Results: 92.3% accuracy, −35% false positives, +60% efficiency; 100+ atomic capabilities online; hundreds of thousands of QPS.

Tencent — WeMM-Embedding Multimodal Embedding Model

• Background: General multimodal embeddings unifying text, image and video for cross-business retrieval inside the WeChat ecosystem.
• Approach: Qwen3-VL backbone; Deep Fusion of multi-level semantics; deduplicated InfoNCE; hierarchical sampler for arbitrary modality mixing.
• Results: MMEB-V2 0.7523 (open-source #1), UVRB 0.686 (open-source #1), +2.6% on Chinese image-text retrieval.

Skills

• LLM Development — SFT / GRPO / PPO, distillation, deployment (vLLM, HuggingFace)
• Multimodal Models — VLM fine-tuning, model merging (SLERP / TIES / DARE), adversarial data synthesis
• Agents & RAG — agent workflow orchestration, retrieval, knowledge bases
• Frameworks & Tools — Python, PyTorch, HuggingFace, FAISS, Linux, Docker

Talks

Teaching

Honors & Awards

• National Encouragement Scholarship
• People’s Scholarship, 2nd Class
• Silver Medal, FLTRP English Competition (Municipal)

Languages

• Chinese (native), English (TOEFL 101 · GRE 327), Japanese (JLPT N2)