DENSE distills evidence from agent execution trajectories into nested shortcut trees, consolidating redundant attempts, identifying resolved subtasks, and retaining useful steps alongside remaining requirements. This structured feedback enables agents to refine their approach on subsequent attempts without needing external rewards or correctness labels.
On Terminal-Bench 2.1, DENSE achieved the highest strict pass rate among feedback methods without external supervision, improving over initial attempts by 7.12 to 15.64 percentage points while using 19.0 to 43.6 percent fewer agent tokens during retries. Across four different agent models, it consistently outperformed self-reflection on hard tasks in cumulative pass rate over multiple feedback iterations, demonstrating potential for continual self-improvement.