GDPevo: Evaluating Agent Self-Evolution on Real Business Tasks

2026-08-04

AI Summary

Existing benchmarks for evaluating agent self-evolution suffer from narrow task coverage, flawed training/test designs, and data contamination. This work proposes GDPevo—the first automated evaluation benchmark tailored to high-value enterprise domains such as CRM, ERP, and finance. GDPevo dynamically recombines business logic between training and test tasks through workflow decomposition, atomic rule assignment, and rule mixing mechanisms, ensuring that performance gains are attributable solely to the evolution process. A fully automated data pipeline is introduced to eliminate contamination risks. Experiments across 120 tasks demonstrate that four representative agents achieve up to a 16.44 percentage point improvement in held-out test accuracy after self-evolution, yet their performance remains substantially below the theoretical upper bound of 91.6%.

Abstract

Agent self-evolution updates an agent's persistent state from prior experience and reuses it to solve related tasks more effectively. Evaluating self-evolution is difficult: existing benchmarks provide limited coverage of economically valuable task domains, do not always design training and test tasks such that test-time gains can be attributed to training experience, and remain vulnerable to data contamination. We present GDPevo, an evolution-native benchmark grounded in GDP-related enterprise workflows, together with the fully automated data pipeline that generates it. Its core mechanism, rule hybridization, decomposes each enterprise workflow into atomic business rules, distributes subsets of these rules across training tasks, and recombines them in held-out test tasks so that test-time gains are attributable. GDPevo spans CRM, ERP, finance, healthcare, legal, and data-centric workflows. Its V1 release contains 120 tasks in 12 groups, with five training and five held-out test tasks per group. Full automation enables the pipeline to expand the suite to 240 tasks in 24 groups (V2) within two days, providing a practical response to contamination. Using GDPevo, we evaluate four agents, each comprising a harness and a model, under four supervision types. Self-evolution consistently improves held-out accuracy by up to 16.44 percentage points. But the best evolved agents remain far below the fully informed oracle ceiling of 91.6%, indicating that the self-evolution ability of current agents remains far from fully realized. We publicly release the pipeline, benchmark, and full evaluation results at https://github.com/Prism-Shadow/GDPevo.

Problem

• agent self-evolution
• benchmark
• data contamination
• enterprise workflows
• task generalization

Innovation

• agent self-evolution
• rule hybridization
• automated benchmark generation
• enterprise workflows
• attributable generalization