Trajectory2Gym lifts tool-use benchmarks 16.9-29.8 points

"Successful tool-use trajectories are not necessarily useful supervision." That is the opening line of a new arXiv submission, and it swings directly at how most teams currently build data for tool-using agents.

The paper argues that even trajectories ending in success can "underrepresent multi-step dependencies or contain actions that are unjustified by the information visible to the agent at decision time." Its fix, called Trajectory2Gym, "constructs executable interactions around under-supported compositional structures and derives supervision only from grounded decisions in their realized executions." Instead of harvesting what the agent happens to get right, the system builds the interactions that fill known gaps and only trains on actions a learner could have justified from the history visible to it.

Across five Qwen3 scales, the method lifts the equal-weight mean of three tool-use benchmarks (Tau1, Tau2, and BFCL) by 16.9 to 29.8 percentage points. "The gains are strongest in stateful and multi-turn interaction, where later decisions depend on earlier observations and feedback," the paper reports.

The abstract publishes no per-benchmark baselines, no comparison against standard rejection sampling, and no numbers on the compute cost of constructing the interactions versus collecting them. The authors reframe the task as "the deliberate construction of learnable interaction experience, rather than the collection of successful trajectories."