Distilling agent behavior into reusable feedback trees improves task performance
DENSE: Distilling Agent Trajectories into Evidence-Grounded Shortcut Trees for Self-Refinement
Artificial Intelligence
Summary
Lots of computer programs called agents can do tasks and leave behind traces showing what they did step-by-step. The authors created a method called DENSE that groups these steps into trees which show progress, problems, and what’s left to do. This helps the agents learn to fix themselves without needing extra labels or help. Experiments showed DENSE helps agents perform tasks better and use fewer steps when trying again.
What this means in practice
- •For robotics software engineers: Use evidence-grounded trees to improve robot task retries without costly human oversight.
- •For automation system developers: Integrate reusable feedback from past executions to boost performance of AI-driven workflow agents.
Authors
Siyuan Liu, Fan Yu, Dongyu Ru, Yizhu Liu, Yifan Yang, Xuezhi Cao, Xunliang Cai, Yixin Cao
Abstract
Online agent deployments produce abundant execution traces, while task-specific verification and expert annotation are costly to scale. We study how to distill these traces into reusable feedback without post-hoc outcome labels, drawing on their evidence of local progress, recovery, and unfinished requirements. We introduce DENSE (Distilling Evidence from Nested Subtask Executions), which organizes this evidence into evidence-grounded nested shortcut trees. DENSE compresses redundant attempts, reconciles issues across levels using recovery evidence, and summarizes completed branches while expanding unresolved ones, linking reusable progress to remaining obligations. We introduce REFIT, a source-paired protocol comparing feedback from shared initial trajectories under post-hoc outcome blindness, with environments and model contexts reset for fresh attempts at the same tasks. On Terminal-Bench 2.1, DENSE achieves the highest strict pass rate among tested non-privileged feedback methods across four recipient models. Relative to initial executions, strict pass rate improves by 7.12-15.64 pp, with 19.0-43.6% fewer observed recipient tokens in reruns. GPT-5.5 ablations support combining nested subtask analysis with shortcut construction and issue reconciliation. These findings point toward agent self-refinement through evidence-grounded trajectory reuse with less reliance on external supervision.