CompoWorld: Compositional Environment Scaling for General Agents
Xiao-Wen Yang, Weiyi Xu, Wen Da, Hang Xu, Canwei Li, Hong-Jie You, Pusen Dong, Yucheng Zeng and 4 others
AllSpark Team
Published September 27, 2026 · on the feed September 29, 2026 · ▲ 38 · arXiv 2609.33665
The abstract
Automatically generated environments provide a scalable source of interaction data for training general agents. However, existing approaches mainly generate tasks within a single environment, while real-world workflows require agents to connect information and actions across multiple services. We introduce Compositional Environment Scaling (CompoWorld), which expands the task space by composing a finite library of reusable services. Coding agents turn tool specifications into verified services with typed states and shared interfaces, while a world model handles tools that cannot be reliably implemented. A random-walk procedure connects services through dependency graphs, enabling the generation and verification of tasks that require information to flow across services. Verified trajectories support supervised fine-tuning (SFT), while our Completion-Focused Rubric Reward guides reinforcement learning (RL) toward full task completion by emphasizing criteria with lower pass rates within each rollout group. We construct 448 services exposing 10,130 tools and use 3K SFT trajectories and 1K RL tasks to train Qwen3.6-35B-A3B. Experimental results show that CompoWorld improves on its backbone by 9.17 points on average across eight benchmarks. On AutomationBench, it surpasses frontier models such as Claude Opus 4.6 and leads all compared agent-specialized 35B-A3B models.
The Brief
2:27A short podcast summary of the paper. Written and voiced by AI.
Transcript
CompoWorld. Compositional Environment Scaling for General Agents. By Xiao-Wen Yang and eleven others, from the AllSpark Team. Published on the twenty-seventh of September. Its code repository is a placeholder: a readme, and nothing to run yet.
Environments generated automatically are a scalable way to get interaction data for training general agents. Most existing approaches, though, generate tasks inside one environment at a time. Real workflows rarely look like that. They ask an agent to carry information and actions across several services, and a single-environment generator never produces that kind of task. An agent trained only on isolated environments never practises the handoff between them, which is exactly where real workflows ask the most.
Compositional Environment Scaling, or CompoWorld, grows the task space by composition. It starts from a finite library of reusable services and builds tasks by connecting them, so the number of possible tasks grows with the ways services can be combined, not only with the number of services.
Building the library falls to coding agents, which turn tool specifications into verified services with typed states and shared interfaces. Tools that cannot be implemented reliably are handled by a world model instead. To make tasks, a random-walk procedure links services through dependency graphs, and this lets the system generate and verify tasks in which information has to flow from one service to another. Verified trajectories feed supervised fine-tuning. For reinforcement learning, the authors add a Completion-Focused Rubric Reward, which pushes the agent toward finishing the whole task by emphasizing the criteria with lower pass rates within each rollout group.
The library holds four hundred and forty-eight services exposing ten thousand one hundred and thirty tools. With three thousand fine-tuning trajectories and one thousand reinforcement learning tasks, the authors train Qwen3.6-35B-A3B. Averaged across eight benchmarks, CompoWorld improves on that backbone by nine point one seven points. On AutomationBench it surpasses frontier models such as Claude Opus 4.6, and it leads every agent-specialized model of the same size that it was compared with.
Picture a set of building blocks rather than a set of finished rooms. Each service is built and verified once, and the variety comes from snapping them together along their dependencies. That is the bet CompoWorld makes: that cross-service tasks, the kind real work is made of, are best produced by composing a modest library rather than by generating each world from scratch.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Real work crosses apps. Generated training tasks usually do not. That is the wall.
You are training a general agent on automatically generated environments. Each generator builds tasks inside one app: an email client here, a database there.
The road ahead
- 1 · The problem ~7 min Agents trained on generated environments only practise tasks inside one app, but real work carries information from one app to the next.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Build apps as reusable services, then link several into one task so information must flow between them.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Build verified services, link them by random walk, generate and check tasks, then fine-tune and run reinforcement learning.
- 6 · The experiment ~6 min You’ll guess: Averaged across eight agent benchmarks, how many points did CompoWorld's composed-environment training add to its Qwen3.6-35B-A3B backbone?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.