GDPevo is an evolution-native benchmark for agent self-evolution, grounded in GDP-related enterprise workflows. It uses rule hybridization to decompose workflows into atomic business rules, distributing subsets across training tasks and recombining them in held-out test tasks to ensure test-time gains are attributable. The V1 release contains 120 tasks in 12 groups, with five training and five held-out test tasks per group. The data pipeline is fully automated, enabling expansion to 240 tasks in 24 groups (V2) within two days. The benchmark spans CRM, ERP, finance, healthcare, legal, and data-centric workflows. Four agents were evaluated using GDPevo.
GDPevo is a new benchmark for evaluating agent self-evolution on real business tasks. It addresses limitations of existing benchmarks by providing broad coverage of economically valuable domains, ensuring test-time gains are attributable to training experience, and mitigating data contamination through automated expansion. The benchmark is built from GDP-related enterprise workflows and uses a rule hybridization mechanism. Its V1 release includes 120 tasks across 12 groups, with plans for rapid expansion to 240 tasks. The benchmark covers CRM, ERP, finance, healthcare, legal, and data-centric workflows. Four agents were evaluated, though specific results are not detailed in the evidence.
The core technical innovation is rule hybridization, which decomposes enterprise workflows into atomic business rules, distributes subsets across training tasks, and recombines them in held-out test tasks. This design ensures that performance improvements on test tasks can be attributed to the agent's self-evolution from training experience, rather than task overlap or memorization. The fully automated data pipeline allows rapid expansion of the benchmark, providing a practical defense against data contamination. The benchmark's grounding in real GDP-related workflows increases its ecological validity for evaluating agent self-evolution in economically significant domains.
GDPevo targets a gap in evaluating AI agents for enterprise workflows, which are critical for economic productivity. By focusing on self-evolution—the ability of agents to improve from experience—the benchmark aligns with industry needs for adaptive, continuously improving AI systems in CRM, ERP, finance, healthcare, and legal domains. The automated expansion capability addresses the persistent challenge of benchmark contamination, making it more suitable for long-term use in enterprise AI development. The benchmark's design may influence how companies evaluate and deploy self-improving agents in real business processes.
GDPevo provides a standardized, scalable way to measure and compare the self-evolution capabilities of AI agents in enterprise contexts. This can help businesses select and refine agents that improve autonomously on tasks like customer relationship management, financial analysis, or legal document processing, potentially reducing manual retraining costs and increasing operational efficiency. The benchmark's focus on attributable gains from experience supports more reliable ROI assessments for self-evolving AI deployments.
Next signals to watch include the release of GDPevo V2 with 240 tasks, publication of detailed evaluation results for the four agents tested, and adoption of the benchmark by AI research labs and enterprise AI teams. The automated pipeline could be extended to generate tasks in additional domains or languages. If the benchmark gains traction, it may spur development of new self-evolution algorithms tailored to enterprise workflows. Potential challenges include ensuring the benchmark's tasks remain representative of real-world complexity and avoiding overfitting to the rule hybridization structure.