Reproduce “Which the Eye Fears: Writing with Read-Blindness Explains Massive Activations in Transformers” 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
Swagatam Mukhopadhyay, Vishal Vivek Saley, Vraj Parikh, Mausam
The authors explain why massive residual-stream coordinates persist in Transformers by measuring how blocks read and write them. Across four models, massive coordinates show weak responses in several input projections while write pathways remain open. In Llama 1.28B, FFN read-blindness separates eventual massive coordinates at 2.3k training steps, before amplifier gain at 4.5k; this ordering does not establish causation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.