DENSE Distills Agent Trajectories into Evidence-Grounded Shortcut Trees

Summary

Online agent deployments generate many execution traces, but task-specific verification and expert annotation are costly to scale. The paper introduces DENSE, or Distilling Evidence from Nested Subtask Executions, to convert those traces into reusable feedback without relying on post-hoc outcome labels. It organizes evidence of local progress, recovery, and unfinished requirements into nested shortcut trees. The method compresses redundant attempts, reconciles issues across levels using recovery evidence, summarizes completed branches, and expands unresolved ones so reusable progress remains linked to outstanding obligations. The authors also introduce REFIT, a source-paired evaluation protocol that compares feedback derived from shared initial trajectories while keeping outcome labels hidden and resetting environments and model contexts for fresh attempts. On Terminal-Bench 2.1, DENSE records the highest strict pass rate among the tested non-privileged feedback methods across four recipient models. Compared with initial executions, strict pass rate rises by 7.12-15.64 percentage points, while observed recipient tokens in reruns fall by 19.0-43.6%. Ablations with GPT-5.5 support combining nested subtask analysis, shortcut construction, and issue reconciliation. The results suggest that evidence-grounded reuse of agent trajectories can support self-refinement with less reliance on external supervision.