Reproduce “Line-Coupled Language Model” 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
Shiyuan Li, Shaorong Zhang, Zhaorui Yang, Qian Zhang, Greg Ver Steeg, Bingyao Li
LCLM generates next tokens for text lines in one forward pass, with causal attention and next-token training. In code and prose tests, cross-line losses changed less when targets were revealed than same-line losses. Matched 881M models reached 2.94 versus 1.00 content tokens per pass, with 0.728 versus 0.713 BPB; quality was worse at line starts, and no wall-clock speedup was measured.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.