Tag
The paper introduces DENSE, a method for distilling AI agent execution traces into evidence-grounded shortcut trees for self-refinement, achieving improved performance on Terminal-Bench without post-hoc outcome labels.