What's this about?
Multi-turn tool-calling is essential for autonomous agents, and progress is limited by training on full-length trajectory imitation. The paper's observation: when a task has multiple order-independent sub-goals, the optimal solution is not a single sequence, so imitating one recorded sequence teaches an arbitrary ordering as if it were necessary. DART-SD uses diamond-topology-aware retrieval and tuning for self-distillation. 72 upvotes.
What is a "diamond topology"?
Picture the task graph: a start, an end, and in between several sub-goals that can be done in any order. Drawn out, the paths from start to finish spread apart and rejoin — a diamond. A recorded trajectory is one path through it. Training on that path teaches the model that A comes before B, when the truth is that A and B are independent. Recognising the diamond means training on the *structure* rather than one traversal of it.
Why does this matter to you?
- It joins SWE-Prime and the good-agentic-data paper in this month's run of "your training data encodes accidents as requirements."
- Order-independent sub-goals describe most real tool-using work, including scanning: check these five things, in whatever order.
- 72 upvotes for a fairly technical training paper suggests the multi-turn tool-calling problem is widely felt.