A tool-using agent can know how to call each function and still fail at the workflow: passing outputs forward, handling state changes, or recovering from errors. The key insight here is to scale training from observed relationships between real tool calls, rather than inventing tools and tasks around synthetic databases.
AI Agent Papers2026
ToolSpace: Scaling Tool-Centric Environment Synthesis for Long-Horizon Agents
Builds tool-composition environments from real execution traces and generates verifiable long-horizon agent tasks. Results show substantial gains in models’ ability to coordinate tools across environments.
Introduction
Information
- Websitearxiv.org
- OrganizationsRenmin University of China, Fudan University, Shanghai Jiao Tong University
- AuthorsGuanting Dong, Zhichao Sun, Xinyu Yang, Rui Wang, Xin Deng, Jiajie Jin, Zhixiong Yang, Junhao Huang, Han Han, Mingyu Ma …
- Published date2026/11/03
More Items
AI Agent Papers2026
Trains agent models with two supervised fine-tuning stages followed by reinforcement learning, and releases the data-generation, verification, and training pipeline. The recipe skips dedicated mid-training, DPO, and on-policy distillation.
Studies how to route tasks to suitable models or policies across general-purpose AI systems while balancing quality and compute cost. Evaluations across 25 benchmarks and over 374,000 runs quantify the resulting capability and efficiency gains.
AI Agent Papers2026
Trains computer-use agents on verified trajectories from long interactions, connecting task generation, execution, evaluation, and policy improvement in a feedback loop. Reports Nano, Mini, and Pro models across desktop and web benchmarks.