M.S. student at Tsinghua University, based in Shenzhen. My current focus is LLM post-training, reinforcement learning, and tool-using agents, with related work in multimodal learning and speech/audio processing.
I build small, inspectable training and evaluation systems: define the task, trace the learning signal, test the implementation, and document what the evidence supports.
Academic homepage · Email · X
| Project | What to explore | Start here |
|---|---|---|
| 5G Diagnostic Agent | Multi-turn tool use, evidence-gated diagnosis, LoRA-SFT/GRPO training, and evaluation | Learning loop · Synthetic-data experiment |
| BLM Multimodal Audit | Visual representation learning, distributed contrastive training, reviewed data production, and GRPO | Data flywheel · Distributed experiments |
| RL From Scratch | Learning implementations and mathematical notes from Bellman equations to PPO, DPO, and GRPO | Post-training chapter |
| Agent the Hard Way | Go exercises on tool loops, permissions, context management, and memory, following Leihb's course | Exercises and verification |
| HearWeave | Microphone-array simulation, beamforming, localization, and reproducible spatial-audio baselines | Tutorial |
| Career Pro AI | A resume-analysis and mock-interview application with a server-side model API and a demo mode | Architecture and setup |
The research repositories distinguish synthetic experiments, implementation checks, and real-world validation. Training configurations and successful tool calls alone are not evidence of model improvement; the linked reports include limitations and unsuccessful experiments.
- Post-training: policy optimization, preference learning, reward design, and training diagnostics.
- Agents: useful tool interactions, synthetic task data, context management, and evaluation.
- Multimodal and audio learning: connecting representation learning with signal processing and distributed computation.
For broader projects and learning resources: JobNebula, math foundations, and my learning hub.