Reproduce “The Transient Nature of Emergent In-Context Learning 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
Aaditya K. Singh, Stephanie C. Y. Chan, Ted Moskovitz, Erin Grant, Andrew M. Saxe, Felix Hill
The authors study whether in-context learning can persist when models also can learn class mappings in their weights. In controlled synthetic exemplar-label classification, they separately evaluate prompt-based matching and stored mappings. In a default 12-layer model trained for 5×10⁷ steps, prompt-based accuracy rises then declines while training loss falls; this suggests, but does not establish, a shift toward in-weights learning.
Key insights
- In-context learning in Transformers is not asymptotically stable; it emerges, peaks, and then decays or disappears during prolonged training, even as overall model loss continues to improve.
- ICL transience is likely due to competition with In-Weights Learning (IWL) circuits, with IWL being asymptotically preferred by the optimization process.
- L2 regularization can stabilize ICL, and certain data characteristics, such as a large number of classes or a moderate Zipfian distribution, can delay its transience.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.