Researchers have proposed a new framework called DART-SD to improve the performance of large language models in tasks that require multi-turn tool-calling. The current approach to this problem relies on full-length trajectory imitation, which can lead to a 'topological collapse' and degrade policy diversity. DART-SD addresses this issue by modeling the execution process as an Interaction-State Transition Graph (ISTG) and identifying critical topological breakpoints. This allows for localized correction and progressive self-distillation, leading to better performance on complex multi-turn tool-calling benchmarks. --- Why it matters: This matters because it can improve the ability of AI agents to perform tasks that require multiple steps and tool usage, such as assembly or repair tasks. By addressing the limitations of current approaches, DART-SD has the potential to enable more efficient and effective autonomous systems. Source: https://arxiv.org/abs/2608.18524