Found first: a primary source the press has not covered yet.
Researchers from ByteDance and the University of Science and Technology of China have published DART-SD, a training framework targeting a structural failure in imitation learning for multi-turn tool-calling agents. Standard fine-tuning penalizes valid alternative orderings of sub-goals, collapsing policy diversity. DART-SD identifies the exact points where trajectories break from valid topology and applies gradient loss only to the recovery steps that follow, leaving correct prefixes untouched.
What the source says
Authors Hangrui Xu, Jiarui Wang, Yang Yang, Chuanbo Zhu, Fangda Chen, Ziqi Wu, Jingming Cai, and Yan Song (ByteDance and USTC) model multi-turn execution as a converging Interaction-State Transition Graph whose topology preserves the diamond-lattice structure of order-independent sub-goals. The method identifies Critical Topological Breakpoints where trajectories diverge and restricts training loss to the generated recovery steps after each breakpoint only. Progressive self-distillation iterates this over five rounds: on a Qwen3-8B backbone, Solve-F1 on the FTRL benchmark rises from 40.37 at iteration 1 to 45.66 at iteration 5, while average tool calls drop from 4.23 to 3.55. An ablation study tracks the full progression from base Qwen3-8B at 23.48 Solve-F1 to the complete method at 45.66. Experiments span five benchmarks: FTRL, BFCL Multi-Turn, ToolHop, τ-bench, and RoTBench, using Qwen3.6-27B and GLM-5.2 as teacher models.
Why it matters
Most agentic training pipelines fine-tune on full recorded trajectories, which treats every deviant path as an error even when it reaches the correct goal through a different valid ordering. For tasks with order-independent sub-goals, the combinatorial space of valid orderings is large, so standard imitation learning penalizes a substantial fraction of correct solutions. DART-SD makes the task topology explicit and confines the training signal to recovery behavior, not the valid steps that preceded the failure. The efficiency gain is a concrete signal: fewer tool calls alongside higher solve rates suggests the policy is learning tighter recovery rather than just longer correct paths. The paper reports general capability scores averaging 49.89 against a 43.92 baseline across IFEval, AIME24, AIME25, and MMLU, indicating no degradation from the selective supervision approach.