Papers.Data.Code
Paper

Trajectory2Gym: Decision-Grounded Compositional Interaction Synthesis for Tool-Use Agents

Yihao Hu, Zhihao Wen, Yitong Lin, Siyuan Lu, Pan Wang, Wei Wu, Jiang-Ming Yang, Tao Lin

Constructs native tool-use interactions around missing multi-step structures, then trains only on actions certified by visible context, execution, and workflow order.

arXiv ↗Oct 24, 2026Code ↗

  • nlp
  • llm
  • tool use
  • supervised fine-tuning
  • data synthesis
Two-stage pipeline diagram showing method: interaction construction and structural-support learning
Two-stage pipeline diagram showing method: interaction construction and structural-support learning

What it is

Builds tool-use training data by synthesizing under-covered interaction paths and supervising only actions justified by the agent-visible history. Instead of imitating successful trajectories, it constructs executable interactions for under-supported multi-step paths and masks supervision to only decision-grounded actions from visible history.

Key results

  • +16.9 to +29.8 pts on equal-weight Tau1/Tau2/BFCL mean across 5 scales
  • Qwen3-4B + Trajectory2Gym gets 69.17 on Tau2 vs 28.33 for Qwen3-30B-A3B init
  • Qwen3-8B + Trajectory2Gym reaches 74.17 Tau2, 28+ pts above all re-evaluated checkpoints

Shows that for tool-use agents, learnable multi-step supervision matters more than simply collecting successful trajectories.

One useful ML email a week.

More from Oct 26 – Nov 01

The whole week →

Picked Oct 30

Picked Oct 29

  • Dataset
    crtproduction/name-that-llm-open-pool

    Unlike most attribution datasets, it provides 254,561 de-duplicated open passages labeled across 18 LLMs with provenance and leakage flags.

    KaggleOct 28, 2026
    →
  • Repo
    google-deepmind/detectorscope

    Unlike black-box AI text detectors, it exposes named feature sums and token-level evidence while preserving accuracy above 0.7 interpretability on COLING and RAID.

    GitHubOct 24, 2026
    →

Picked Oct 28

  • Dataset
    baidu/MarathonDocBench

    Makes hundred-page parsing measurable with cross-page structure and 8 document-level metrics, exposing failures of page-wise and general VLM parsers.

    Hugging FaceOct 26, 2026
    →

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.