Hugging Face researchers have unveiled DART-SD (Diamond-topology Aware Retrieval and Tuning for Self-Distillation), a novel framework designed to enhance multi-turn tool-calling agents. The approach addresses a fundamental limitation in current training methods that treat tool-use trajectories as linear sequences, which can inadvertently penalize valid alternative explorations.
Traditional supervised fine-tuning (SFT) and reinforcement learning (RL) methods force agents to imitate full-length trajectories. However, for tasks with multiple order-independent sub-goals, the optimal solution space forms a vast combinatorial diamond lattice. Forcing this rich topology into monolithic trajectories causes a 'topological collapse,' reducing policy diversity and hindering learning.
DART-SD shifts the paradigm by modeling the execution process as an Interaction-State Transition Graph (ISTG), capturing the inherent diamond topology of successful and failed exploratory paths. During autonomous rollouts, the framework identifies the Critical Topological Breakpoint (CTB) — the point where reasoning goes wrong — and retrieves success-supported recovery references.
The framework then applies a progressive self-distillation paradigm through CTB-guided localized supervision. Training loss is calculated exclusively on the generated recovery steps, protecting the valid reasoning prefix from destructive gradient updates. This ensures that agents learn precisely where errors occur and how to correct them, without disrupting already-correct reasoning.
Experiments across five tool-use benchmarks and two model scales show that DART-SD consistently outperforms strong SFT and RL baselines. Notably, the Qwen3-8B student model even surpasses its much larger teacher on several benchmarks, demonstrating the effectiveness of topology-guided localized correction.