DENSE Turns Raw Agent Traces Into Reusable Feedback Without Any Outcome Labels

arXiv 2609.21423·medium signalXBlueskyLinkedInCopy link

DENSE distills online agent execution traces into evidence-grounded nested shortcut trees, compressing redundant attempts, reconciling issues across levels using recovery evidence, and summarizing completed branches while expanding unresolved ones, all without post-hoc outcome labels or expert annotation. Evaluated under REFIT, a source-paired protocol that resets environments and model contexts for fresh attempts from shared initial trajectories, DENSE gets the highest strict pass rate among tested non-privileged feedback methods across four recipient models on Terminal-Bench 2.1. Strict pass rate improves 7.12 to 15.64 percentage points over the initial executions while the reruns consume 19.0-43.6% fewer recipient tokens.