DART-SD trains tool-calling agents by patching only the steps where the trajectory went wrong

arXiv·medium signalXBlueskyLinkedInCopy link

An August 19 arXiv paper argues that forcing diverse solution paths into single training trajectories causes topological collapse, and replaces full-trajectory learning with topology-guided localized correction. It models execution as an Interaction-State Transition Graph, identifies Critical Topological Breakpoints where paths diverge, and applies progressive self-distillation with gradient updates restricted to recovery steps so already-correct reasoning is left untouched. The authors report it significantly outperforms full-trajectory baselines on complex multi-turn tool-calling benchmarks, though the abstract does not give numbers.