- The problem
- Agents that work for many steps on complicated tasks are hard to train: creating trustworthy practice tasks takes substantial effort, and a task that is graded only at the end gives no useful feedback when the agent does not finish. This is especially limiting for less capable open models.
- The idea
- The paper first builds a complete, correct hidden version of a task, then creates the agent’s challenge by rendering and deliberately damaging parts of it. For example, it can start with a correct data warehouse, introduce messy or inconsistent files, and use the hidden original to check both the final repair and intermediate progress. That progress becomes a reward signal at each step, rather than feedback arriving only after full completion.
- What it got
- The authors provide 1,170 verified practice environments spanning seven areas. A model trained with their approach reached 0.28 Pass@5 on the computing-and-math exam split, matching GPT-6 Astra and Claude Opus 5.