Trajectory2Gym: A new method to train tool-use agents
In brief
- Trajectory2Gym targets multi-step dependencies in tool-use agent trajectories.
- The method generates executable interactions with supervision grounded in visible history.
- Improves Qwen3 performance by 16.9 to 29.8 points across three benchmarks.
Researchers introduce Trajectory2Gym, a method to enhance training for agents capable of using tools. According to the announcement, successful trajectories alone are insufficient as supervision, as they may miss multi-step dependencies or include actions not justified by the agent's visible history.
Trajectory2Gym builds executable interactions by targeting underrepresented compositional structures and deriving supervision only from decisions grounded in the accessible history. Tested on five scales of the Qwen3 model, the method improves performance on three benchmarks (Tau1, Tau2, BFCL) by 16.9 to 29.8 points on average, with stronger gains in stateful and multi-turn interactions.
This summary gives you the essentials. The full article, its figures and quotes are on Hugging Face Daily Papers.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error