Reproduce “Attention Sinks Induce Gradient Sinks: Massive Activations as Gradient Regulators 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
Yihong Chen, Zhouchen Lin, Quanming Yao
The authors propose V-scale to attenuate backward signals along attention’s value path. They interpret massive activations as responses to gradient pressure at attention sinks. In a 0.3B model, median gradient-sink ratio falls from 40.3 to 5.6, but sinks remain. In a separate 1B multi-key retrieval test, accuracy rises from 48.32% to 62.00%. This establishes neither general performance gains nor a sole cause.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.