Trajectory2Gym builds decision-grounded supervision for tool-use agents
Original title: Trajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents
Across five Qwen3 scales it lifts the equal-weight Tau1, Tau2 and BFCL mean by 16.9 to 29.8 points.
Trajectory2Gym argues that successful tool-use traces can miss multi-step dependencies or include actions not justified by what the agent could see. It constructs executable interactions around under-supported compositional structures and keeps supervision only from decisions grounded in the realized, learner-visible history. Structural-support selection allocates experience, and parent-normalized supervised fine-tuning treats the semantic interaction as the training-weight unit. Tested across five Qwen3 scales, it improves Tau1, Tau2 and BFCL at every scale, with the largest gains in stateful, multi-turn settings.
- The equal-weight three-benchmark mean rises by 16.9 to 29.8 percentage points.
- Later decisions in the strongest gains depend on earlier observations and feedback.
- The method rejects collecting successful trajectories as sufficient supervision.
From the original article
1 min read at huggingface.co
Successful tool-use trajectories are not necessarily useful supervision. Even successful interactions may underrepresent multi-step dependencies or contain actions that are unjustified by the information visible to the agent at decision time.
By Yihao Hu, Zhihao Wen, Yitong Lin · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.