Reproduce “Rethinking Sequence Compression with Test-time Regression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sihan Yang, Jie Huang, Hanxuan Li, Yifan Guo, Weigao Sun, Yiran Zhong, Weiyang Liu
TTR fits input-specific summaries from token relationships for sparse long-context attention and multimodal vision models. It learns a target projection and regularizer, then solves coefficients anew for each input while keeping parameters fixed. In language retrieval, convolutional TTR reaches 0.5144 RULER accuracy versus 0.3920 for attention-weighted pooling; in vision, 3×3 merging scores 54.85 versus 54.93 for 2×2 MLP, using 55.6% fewer tokens.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.