I am an undergraduate in Computer Science at Tsinghua University. Since June 2026, I have been conducting research with Prof. Hao Zhang at UC San Diego’s Hao AI Lab, working on efficient language model architectures and parallel decoding.
Previously, I worked with Prof. Beidi Chen at Carnegie Mellon University and Prof. Christopher De Sa at Cornell University. At Tsinghua, I worked with Prof. Jianfei Chen and Prof. Youyou Lu.
My research focuses on making language models more capable under practical compute, latency, and memory constraints. I work across model architecture, training algorithms, and GPU systems.
Reuses completed cross-trajectory representations to improve training stability and depth scaling while preserving parallel execution. With eight loops, improves average downstream accuracy by 0.84 percentage points over vanilla PLT with 0.3% additional decoding time.
A dynamic sparse-attention schedule controls actor-policy mismatch for stable long-context reinforcement learning, delivering up to 2.4× rollout speedup.
YAQA targets full-model output error during post-training quantization. Kronecker-factored Hessian sketches reduce KL divergence by approximately 30% over GPTQ/LDLQ without adding inference overhead.