Reproduce “Beyond the Training Horizon: Mechanisms and Limits of Length Generalization in Looped 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
Jia Liang, Xi Jin, Liangming Pan
The paper analyzes how two looped Transformer configurations extend computation beyond training lengths. In controlled composition and knowledge-graph tasks, models receive only final-answer supervision. Analysis finds MR-Loop advances relation selection while DR-Loop propagates a state frontier; both require usable computational states. At unseen relation index 14, an intervention raised MR-Loop correct-hop accuracy from 9.4% to 93.8%. Findings concern small synthetic models, not larger or naturalistic systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.