Publications

Rethinking Sequence Compression with Test-time Regression
Sihan Yang, Jie Huang, Hanxuan Li, Yifan Guo, Weigao Sun, Yiran Zhong, Weiyang Liu†
arXiv 2026
tl;dr: TTR-Compressor uses test-time regression to compress tokens. For sparse attention, it improves retrieval accuracy; for visual token merging, it matches the standard MLP merger's performance with 55.6% fewer tokens.

Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation
Kexuan Shi*, Hanxuan Li*, Zeju Qiu, Yandong Wen, Simon Buchholz, Weiyang Liu
NeurIPS 2026
tl;dr: A spectrum-preserving optimizer that updates each weight matrix via left and right orthogonal transformations, keeping singular values fixed while modulating geometry; competitive for LLM pretraining and finetuning.

DVD-Quant: Data-free Video Diffusion Transformers Quantization
Zhiteng Li*, Hanxuan Li*, Junyi Wu, Kai Liu, Linghe Kong, Guihai Chen, Yulun Zhang, Xiaokang Yang
ICLR 2026
tl;dr: We develop the first data-free quantization framework that achieves W4A4 PTQ for Video DiTs with 2× speedup while maintaining visual fidelity.