Undergraduate researcher interested in efficient algorithms and systems for large language models.
I am currently pursuing a B.S. in Data Science and Big Data Technology at Yuanpei College, Peking University. My work focuses on algorithm-system co-design for efficient LLM post-training, reinforcement learning, and emerging model architectures that make large models more efficient, reliable, and deployable.
Studying efficiency bottlenecks in LLM reinforcement learning, including long-tail rollout lengths, straggler effects, and length explosion during training. This work includes DARTS, a distribution-aware rollout shaping system, and QLPO, a length-aware policy optimization method.
Exploring diffusion language models for efficient parallel generation, with attention to the quality-efficiency trade-off, scalable training, long-block generation, and practical serving efficiency.
Worked on MemClaw-Bench, a capability-centered benchmark for evaluating context management in compressed-context, long-horizon tool-using agents.
Yujie Wang*, Siwei Chen*, Longzan Luo*, Xinyi Liu, Xupeng Miao, Fangcheng Fu, Bin Cui. Co-first author.
arxiv.org/abs/2605.30859Siwei Chen, Siqi Chen, Xupeng Miao, Bin Cui. First author.
Siwei Chen, Xupeng Miao, Bin Cui. First author.
Yaqi Zhao, Yuanyang Yin, Lin Li, Mingan Lin, Victor Shea-Jay Huang, Siwei Chen, Weipeng Chen, Baoqun Yin, Zenan Zhou, Wentao Zhang.
arxiv.org/abs/2411.16824Retrieval-augmented LLM question-answering framework with source tracing, predictive question generation, and learner state diagnosis.
GitHubA CUDA, C++, and Python implementation covering basic kernels, Python-C bindings, a simple computational graph, and CNN training for image recognition.
GitHubPython, C++, Rust, PyTorch, CUDA