Tuesday, 27 October 2026 · 6 min read
Trajectory2Gym improves tool-use agents by synthesizing and grounding compositional interactions
Trajectory2Gym constructs training data focusing on compositional coverage and decision grounding, improving tool-use agent performance across five Qwen3 model scales by 16.9–29.8 points on three benchmarks, especially in multi-turn, stateful tasks.
Based on the paper Trajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents by Yihao Hu, Zhihao Wen, Yitong Lin, Siyuan Lu and 4 others
Challenges in supervising tool-use agents with compositional interactions
Tool-use agents perform multi-step workflows where later decisions depend on earlier observations and feedback, called compositional interactions. For example, a support agent must verify information before canceling orders, requiring the agent to learn dependencies across steps.
Supervising such agents typically uses successful trajectories, but success alone does not guarantee that all important observation-action dependencies are represented or that each action is justified by the agent's visible history at decision time.
Two key challenges arise: compositional coverage, ensuring training data represent multi-step dependencies adequately, and decision grounding, ensuring each supervised action is supported by information visible to the agent when it acts. Without addressing these, training may imitate shortcuts or unsupported actions, limiting agent performance.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Two key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short