Academic homepage · ORCID · Email · X
I'm an M.S. student at Tsinghua University, based in Shenzhen. I work on LLM post-training and tool-using agents, with a background in multimodal learning and speech/audio processing.
I like turning ideas into systems you can run, inspect, and improve—from learning objectives and agent interactions to training code and experiment reports.
- Generative recommendation: studying unified generation, candidate attribution, and the boundary between model, strategy, ranking, and feedback gains.
- Speech enhancement: studying content-preserving enhancement under nonstationary noise and competing speakers.
- Verifiable post-training: building evaluation loops where rewards, trajectories, and held-out behavior can be inspected independently.
I keep public claims tied to runnable protocols, measured outcomes, and explicit evidence boundaries.
Kimi K3 · Deep Dive & Lab is my focused study of foundation model mechanisms, from KDA memory and AttnRes to MoE routing. I connect source-grounded derivations with runnable NumPy experiments, a pre-training and systems guide, and an agentic post-training lab.
Learning route · Numerical report · Pre-training & systems · Agent Lab
| Agents & post-training | Multimodal learning |
|---|---|
 |
 |
| Multi-turn diagnosis with evidence-guided tools, LoRA-SFT/GRPO, and trajectory-level evaluation. | Visual representation learning, distributed contrastive training, and a reviewed data flywheel. |
| Learning loop · Experiments | Data flywheel · Distributed training |
| Speech & spatial audio | Reinforcement learning |
 |
 |
| Microphone-array simulation, beamforming, localization, and reproducible wearable-audio experiments. | Derivations and NumPy/PyTorch implementations, from Bellman equations to PPO, GRPO, and agentic post-training. |
| Tutorial · BeamBench | Learning path · Agentic post-training |
The common thread is learning from interaction: build an environment, follow the signal through the update, and measure what changes.
Explore more: Go agent exercises · JobNebula · Research training guide · Learning hub