AttnRL

Attention as a Compass – Teaching Reasoning Models to Explore Smarter

AttnRL uses a model's own attention to pick branching points in process-supervised RL, beating GRPO and TreeRL by 1.8 points on math while cutting training time 8%.