Reproduce “Deep Latents, Long Memory: Latent persistence explains temporal correlations in transformer activations” 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
Francesco Cagnetta, Santiago Acevedo
The authors measure how correlations between language-model activations change with token distance, using average dot products between positions. In a controlled hierarchy, known persistent latent variables explain much of the measured pattern; in natural text, deeper layers show slower-decaying correlations, and Gemma 2 9B’s layerwise persistence and integrated autocorrelation correlate at r = 0.93. This supports, but does not prove, a role for persistent features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.