HeyHomees Lab Notes · No. 10 · ant-international · Published October 27, 2026
By HeyHomees Newsroom ·
In 30 seconds
- Creates supervised tool-use data that ensure each action is justified by the agent's visible history.
- Improves the equal-weight mean of three benchmarks by 16.9-29.8 percentage points across five model scales.
What is it?
Trajectory2Gym is a code release that constructs executable tool-use interactions around under-supported compositional structures and derives supervision only from grounded decisions in their realized executions. It includes interaction representation, decision-grounded target construction, structural-support selection, and parent-normalized supervised fine-tuning.
How does it work?
The system selects structural support to allocate experience, then uses parent-normalized supervised fine-tuning to preserve semantic interactions as training weight. It renders interactions via thin native environment adapters and trains with ordinary assistant-token cross-entropy, not an RL objective.
How well does it do?
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. Gains are strongest in stateful and multi-turn interaction.
What does it need to run?
Python 3.10+ (or 3.11 for CI), Linux, PyTorch, and symbolic-link support. CPU tests use the CPU build; GPU training requires a compatible PyTorch build. The repository includes regression tests that run on small synthetic inputs without benchmark credentials.
By the numbers
- Mean benchmark improvement: 16.9-29.8 percentage points
- Public candidate interactions: 34,772
- Dataset size: about 125 MB
- 1.7B checkpoint size: 8.1 GB
- 30B-A3B checkpoint size: 61.1 GB
The honest read
The reported 16.9-29.8 percentage-point gain is self-reported by the authors and reflects improvements on three benchmarks (Tau1, Tau2, BFCL) across five model scales; it does not include comparisons to other methods or external baselines. The checkpoints are provided at the precisions and file sizes listed, but the release does not state inference speed, memory requirements, or performance on real-world tasks beyond the benchmark suite. The code runs on Linux with PyTorch and requires Python 3.10+; GPU performance depends on the user-selected PyTorch build.
What we don't know yet
- Licensing terms for the released code and checkpoints beyond the unspecified platform licence
- Performance on downstream real-world tool-use tasks not covered by the three benchmarks
Sources
- Paper (arXiv abstract) (primary)
- Code, GitHub
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.