DART-SD: Diamond-topology Aware Retrieval and Tuning for Self-Distillation of Multi-Turn Tool-Calling Agents
The paper argues that full-trajectory imitation trains tool agents against valid alternate paths.
DART-SD represents multi-turn tool use as an interaction-state graph, so order-independent sub-goals are not collapsed into one “correct” route. It identifies where a rollout first breaks from recoverable behavior, then retrieves successful recovery references from that point. Training loss is applied only to the recovery steps, leaving the valid prefix untouched. The authors report stronger results than full-trajectory baselines on complex multi-turn tool-calling benchmarks. HF Daily Papers' note
DART-SD represents multi-turn tool use as an interaction-state graph, so order-independent sub-goals are not collapsed into one “correct” route. It identifies where a rollout first breaks from recoverable behavior, then retrieves successful recovery references from that point. Training loss is applied only to the recovery steps, leaving the valid prefix untouched. The authors report stronger results than full-trajectory baselines on complex multi-turn tool-calling benchmarks. HF Daily Papers' note
score 5