A recent research paper from Hugging Face challenges the prevailing trend of scaling up multimodal environment pools for training AI agents. The authors find that simply increasing the number of environments does not always improve performance and can even lead to negative transfer and optimization conflicts.
To address this, they propose two complementary methods: Ability-aware Environment Selection (AES) and Hierarchical Difficulty Curriculum (HDC). AES selects environments that provide broad capability coverage, low redundancy, and reduced conflicts, while HDC organizes training through two difficulty levels: harness weakening and state-scale progression.
Experiments demonstrate that these methods substantially outperform naive environment scaling, leading to better training efficiency and generalization. The work highlights the importance of designing environment distributions along two dimensions: diversity and difficulty structure.