Reproduce “Large Language Models as Nonlinear Dynamical Systems: The Three Clocks of Depth, Training, and Generation” 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
Maor Farid
The authors present a dynamical-systems framework linking token representations across layers, weights during training, and generated tokens fed back into context. They map fifteen concepts to each clock and label evidence as proved, measured, or conjectured. In GPT-2, escape slopes track loop logit gaps, with a median relative difference of 6.9%. Barriers saturate, so escape temperature depends on generation horizon.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.