Hugging Face researchers have released GDPevo, a new benchmark designed to evaluate how AI agents can improve themselves by learning from past experience on real business tasks. The benchmark focuses on enterprise workflows tied to GDP-related domains, including CRM, ERP, finance, healthcare, legal, and data-centric operations.
The core innovation is rule hybridization, which decomposes each workflow into atomic business rules. These rules are distributed across training tasks and recombined in held-out test tasks, ensuring that any test-time performance gains can be directly attributed to the training experience. This design also mitigates data contamination, a common issue in agent evaluation.
The V1 release includes 120 tasks in 12 groups, with five training and five test tasks per group. The fully automated pipeline can expand the suite to 240 tasks in 24 groups (V2) within two days, providing a practical response to contamination.
Evaluating four agents under four supervision types, the researchers found that self-evolution consistently improves held-out accuracy by up to 16.44 percentage points. However, the best evolved agents still fall far short of the fully informed oracle ceiling of 91.6%, indicating that current agents' self-evolution ability is far from fully realized.
The pipeline, benchmark, and full evaluation results are publicly available on GitHub.