Verifiable Environments Are LEGO Bricks: Recursive Composition for Reasoning Generalization

Hao Xiang, Qiaoyu Tang, Le Yu, Yaojie Lu, Xianpei Han et al.|June 10, 2026arXiv

Key Takeaway

You can generate unlimited diverse reasoning training environments by recursively composing smaller, reusable building blocks—reducing manual environment creation from 300 tasks to 50 while maintaining or improving model performance.

Summary

This paper introduces RACES, a framework that treats verifiable environments (structured tasks for training reasoning) as composable building blocks that can be automatically combined.

training reasoning agents

Key Terms

verifiable-environments recursive-composition reinforcement-learning-from-verifiable-rewards reasoning-generalization