Junxi Yin

DRIFT: Difficulty Routing Self-DIstillation with Rhythm-Gated Exploration and Success BuFfer Training

Haisen Luo, Yiwei Liu, Haoning Wang, Dan Liu, Junxi Yin, Haotian Wang, et al.

Enabling large language models to achieve stable self-improvement without external expert supervision remains a central challenge in complex reasoning tasks. Existing self-distillation and reinforcement learning methods lack explicit mechanisms for tracking problem-level learning…Read more