Many-Shot CoT-ICL: Making In-Context Learning Truly Learn

Researchers studied many-shot chain-of-thought in-context learning across reasoning and non-reasoning tasks.

They found that scaling, retrieval, and ordering affect reasoning models differently, suggesting long prompts act like structured curricula.

A new ordering method, Curvilinear Demonstration Selection, improved performance by up to 5.42 points on geometry.

The results show that example selection and ordering matter more for reasoning tasks than simple retrieval alone.