Two papers accepted by NeurIPS 2025.

• When and How Unlabeled Data Provably Improve In-Context Learning
• BREAD:Branched Rollouts from Expert Anchors Bridge SFT & RL for Reasoning