Found first: a primary source the press has not covered yet.
A paper posted to arXiv on October 24, 2026 introduces Trajectory2Gym, a framework that builds tool-use training data by constructing interactions around the compositional structures a model cannot yet handle, rather than collecting trajectories where the agent happened to succeed. Authors Yihao Hu, Zhihao Wen, Yitong Lin, Siyuan Lu, Pan Wang, Wei Wu, Jiang-Ming Yang, and Tao Lin report an equal-weight mean improvement of 16.9 to 29.8 percentage points across Tau1, Tau2, and BFCL over five Qwen3 model scales.
What the source says
The framework identifies "under-supported compositional structures," multi-step observation-action dependencies that successful-trajectory collection tends to miss because the model already handles simpler cases. It builds executable interactions around those gaps and applies parent-normalized supervised fine-tuning, treating the semantic interaction as the unit of training weight and restricting supervision to actions grounded in the learner-visible history at decision time. The five Qwen3 scales tested are 1.7B, 4B, 8B, 14B, and 30B-A3B. On Tau1 and Tau2 combined, per-scale mean gains range from 20.04 to 40.92 points; BFCL gains range from 3.35 to 11.93 points. The post-trained 4B model scored 69.17 on Tau2, against 28.33 for the public 30B-A3B initialization.
Why it matters
Standard tool-use training pipelines filter by success: if the trajectory reached the goal, it enters the training set. This paper argues that success is not the right filter, and that compositional coverage and decision grounding predict learning outcomes better than trajectory completion. A 4B model trained with Trajectory2Gym scored 69.17 on Tau2 against 28.33 for the much larger 30B-A3B public baseline. For teams building or fine-tuning tool-use agents, the practical question shifts from generating more successful rollouts to identifying which multi-step interaction structures are missing from the training distribution.