Reproduce “Stabilizing Extrapolation in Looped Transformers via Learned Stochastic Stopping” 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
Hsun-Yu Kuo, El Mahdi Chayti, Patrik Reizinger, Wieland Brendel, Martin Jaggi
The authors study how training-time stopping depth affects length generalization in Looped Transformers, which repeatedly apply the same block. They compare fixed-depth, sampled-depth, and learned stopping schedules on algorithmic tasks. On binary addition, RL-Halting reaches 45.0% mean oracle OOD accuracy versus 30.9% for K=n±5, but results vary by task and do not guarantee reliable extrapolation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.