About me
I’m a second-year Ph.D. student in Computer Science at the Legal Tech Group of the Technical University of Munich (TUM), advised by Prof. Matthias Grabmair. My research focuses on post-training and alignment for legal large language models, with particular interests in reinforcement learning, on-policy distillation, agentic search and tool use, model evaluation, and reward modeling.
I received my M.Sc. in Robotics, Cognition, Intelligence from TUM in 2025. My master’s thesis, completed in collaboration with Google DeepMind, studied relation-specific neurons in multilingual LLMs and led to a paper at EMNLP 2025 Main. I hold a B.Eng. in Mechanical Engineering from Southeast University, Nanjing, China.
News
Research Experience
Generatives Sprachmodell Justiz (GSJ)
LLM Post-Training · Alignment · Legal NLP Systems
- Develop scalable post-training recipes and distributed pipelines using reinforcement learning and on-policy distillation for legal reasoning and summarization.
- Engineer runtime infrastructure for tool-using legal agents, including context management, sandboxed execution, state machines, and multi-turn workflow integration.
On-Policy Self-Distillation for Legal Reasoning
LLM Post-Training · Knowledge Distillation · Legal Reasoning
- Designed on-policy self-distillation workflows for compact LLMs on the LEXAM reasoning benchmark.
- Studied how privileged information, including gold rationales, statutory hints, and teacher critiques, can improve training stability and out-of-distribution generalization.
Answer-Aware Hint Refinement for RL-Trained Search Agents
Agentic RAG · Reinforcement Learning · Tool Use
- Developed answer-aware, loss-masked hint refinement for multi-hop search agents to address credit assignment and reward sparsity.
- Built distributed RL training infrastructure across an 8xH200 GPU cluster using VeRL, vLLM, GPU-accelerated FAISS, and an asynchronous LLM-judge reward server.
Open Source
Zotero Smart Highlighter
Intelligent Document Assistant · TypeScript · Python
An open-source Zotero 8 plugin that extracts and highlights informative passages in research papers through configurable backends: local BM25 ranking, an on-device neural reranker, or LLM APIs.