1/ New preprint! Reasoning models often require hundreds of task examples and thousands of rollouts to improve on a task. How can they learn more from much less?
Introducing CORE: contrastive self-reflection for rapid, sample-efficient, and interpretable self-improvement 🧵
