Construct interactions
Target under-supported capability paths and execute the resulting tasks in native environments. Make discovery, recovery and cross-step dependencies part of the experience.
TOOL-USE AGENTS / RESEARCHTRAJECTORY → EXPERIENCE → LEARNING
Trajectory2Gym
+16.9–29.8 pp
Three-benchmark mean gain
5 model scales
Qwen3 · 1.7B to 30B-A3B
1 epoch of SFT
Parent-normalized training
Conceptual method illustration
Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents
Code, training scripts, examples, and tests are available on GitHub. ↗
01 / THE IDEA
A successful action.
But a useful lesson?
A tool call can execute perfectly while ignoring the user's latest request. Training should preserve the dependencies that make a decision justified.
FROM THE PAPER'S CONSTRUCTED EXAMPLE
“Actually, just the lamp.”
- 01 The user first requests a larger bottle and a dimmer lamp.
- 02 The catalog exposes valid replacement options.
- 03 The user confirms only the lamp exchange.
Both candidate exchanges can execute. Only one respects the confirmed request.
Exchange the lampKeep the original bottle
KeepExchange both itemsReplace the bottle as well
RejectGrounding checks visible evidence; workflow certification checks the confirmed scope and required order.
Paraphrased schematic from Figure 1; the second candidate is constructed. This interaction illustrates supervision selection.Read the full abstract +
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. We identify two coupled requirements for effective trajectory supervision: compositional coverage, so that important observation–action dependencies are represented, and decision grounding, so that each supervised action is supported by the learner-visible interaction history. We introduce Trajectory2Gym, which constructs executable interactions around under-supported compositional structures and derives supervision only from grounded decisions in their realized executions. Structural-support selection allocates the resulting experience, while parent-normalized supervised fine-tuning preserves the semantic interaction as the unit of training weight. This reframes trajectory generation as the deliberate construction of learnable interaction experience, rather than the collection of successful trajectories. Across five Qwen3 scales, Trajectory2Gym improves Tau1, Tau2, and BFCL at every tested scale, increasing the equal-weight three-benchmark mean by 16.9–29.8 percentage points. The gains are strongest in stateful and multi-turn interaction, where later decisions depend on earlier observations and feedback. Our results suggest that scaling tool-use agents depends not only on collecting more successful trajectories, but on constructing interactions whose dependencies are represented and whose decisions are learnable from the agent's visible history.
02 / THE METHOD
Build. Ground.
Allocate. Learn.
Construct under-supported interactions, certify decisions in their realized executions, and distribute learning across semantic trajectories.
Certify decisions
Check visible evidence, execution validity and workflow completion. Keep authentic errors as context while supervising the grounded recovery that follows.
Allocate support
Select certified trajectories to preserve task breadth and cover capability paths. Balance repeated support with a scale-dependent target.
Normalize by parent
Give each semantic interaction equal training weight, independent of how many examples it renders into. Train with one epoch of supervised fine-tuning.
03 / THE EVIDENCE
Every scale.
All three benchmarks.
Paired comparisons against each Qwen3 initialization, under the same task population, native interface, decoding and scorer.
Stateful interaction · Tau2
Pass@1 (%), equal-weight domain average. Higher is better.
| Qwen3 scale | Initialization | + Trajectory2Gym | Gain (pp) |
|---|---|---|---|
| 1.7B | 23.33 | 37.50 | +14.17 |
| 4B | 19.17 | 69.17 | +50.00 |
| 8B | 25.00 | 74.17 | +49.17 |
| 14B | 25.00 | 80.83 | +55.83 |
| 30B-A3B | 28.33 | 84.17 | +55.84 |
Author-reported manuscript results. Each task is run once. Scale-conditioned training views differ; these gains measure the complete recipe. Tau1 Airline uses a seen-task, transductive setting.
Training weight matters.
Parent normalization delivers the largest mean gain among the four stage-matched 8B ablations, with trajectories, rendered examples and target masks unchanged.
Broader structural coverage.
Structural selection covers more canonical paths than random selection at the same 14,819-parent budget and equal target tokens.
Read the ablation context +
Each matched ablation is a single run. A few-point change is directional, and gains from different stages are not additive. The targeted-expansion pair changes candidate supply without fixing the target-token budget. The support-depth sensitivity experiment changes both support and data volume. See the full paper for protocols and limitations.
04 / THE BEHAVIOR
Use the evidence.
Respect the order.
On the paper's Retail workflow, the address must change while the order is pending, before editing its items. Correct individual calls still need the right sequence.
05 / RESOURCES
Explore the work.
Explore the paper, implementation, training workflow, and project materials.
CITATION
Build on this work.
BibTeX for the project manuscript.
@misc{hu2026trajectory2gym,
title = {Trajectory2Gym: Decision-Grounded
Compositional Interaction Synthesis
for Tool-Use Agents},
author = {Hu, Yihao and Wen, Zhihao and
Lin, Yitong and Lu, Siyuan and Wang, Pan and
Wu, Wei and Yang, Jiang-Ming and Lin, Tao},
year = {2026},
howpublished = {Project manuscript},
url = {https://yihaohu0118.github.io/Trajectory2Gym-page/}
}