Transformers Learn to Implement Multi-step Gradient Descent with Chain of Thought.

Jianhao Huang, Zixuan Wang, Jason D. Lee

Browse the full ICLR paper archive.