Found first: a primary source the press has not covered yet.
A paper posted to arXiv argues that successful tool-use trajectories are not necessarily useful training data, and introduces a method called Trajectory2Gym that constructs interactions around compositional gaps instead. Across five Qwen3 model scales, the approach raises an equal-weight mean across three tool-use benchmarks by 16.9 to 29.8 percentage points. The full details are in the October 24, 2026 arXiv submission.
What the source says
The work, from eight researchers including Yihao Hu, Zhihao Wen, Yitong Lin, and Tao Lin, identifies two requirements their method satisfies: compositional coverage and decision grounding. Trajectory2Gym uses structural-support selection to steer experience toward compositional structures that existing successful trajectories underrepresent, then pairs this with parent-normalized supervised fine-tuning that treats semantically coherent interactions as the training unit. Supervision is restricted to actions grounded in information the model could actually observe at decision time. The benchmarks tested are Tau1, Tau2, and BFCL, with the strongest gains appearing in stateful and multi-turn interaction settings.
Why it matters
Agent training pipelines are typically built around filtering for successful rollouts, on the assumption that more success equals better signal. This paper argues the bottleneck is elsewhere: if training interactions lack the compositional dependencies the agent needs to generalize, collecting more successes does not close the gap. The 16.9 to 29.8 point range holding across five model scales suggests the effect is not confined to a particular model size. For practitioners investing in tool-use data pipelines, the paper reframes the optimization target from success rate toward structural coverage.