GDPevo is a new benchmark for evaluating agent self-evolution on enterprise business tasks, published August 4, 2026, by Xi'an Jiaotong University and collaborating institutions. It covers 240 tasks across 24 groups in CRM, ERP, finance, healthcare, legal, and data analysis workflows. The benchmark includes an automated data pipeline and contamination defense: when existing tasks leak, fresh ones regenerate in two days.
The core mechanism is "rule hybridization." Each enterprise workflow decomposes into atomic business rules. These rules distribute across training tasks, then recombine in new patterns for held-out test tasks. This ensures accuracy gains reflect experience with the training distribution, not memorized answers. The construction pipeline runs three automated stages: scenario discovery, task group synthesis, and quality filtering (six reviewer-agent votes per task group).
Four evaluation modes map to production realities. Base requires no evolution. Self uses training inputs plus environment feedback, no gold answers. Reflect-3 uses judge feedback from training tasks, iterated three times, no gold answers. Fewshot includes gold answers as demonstrations. Ground-truth labels are expensive or unavailable in most deployments.
In self mode — the most realistic production setting — top models gain roughly 8 percentage points. Claude Code with Opus 4.8 improves from 49.11% to 57.37%. Codex with GPT-5.5 moves from 46.72% to 54.99%. Claude Code with GLM-5.2 goes from 47.73% to 55.91%. Self-evolution also reduces spend in most configurations. GPT-5.5 sees a 32.18% cost drop, GLM-5.2 a 10.23% drop. Evolved agents complete tasks in fewer model-response rounds.
Gains scale with supervision. With gold answers, Opus 4.8 reaches 70.90% (+21.79 pp), Opus 4.6 via Panofy reaches 71.47% (+21.07 pp), and GLM-5.2 reaches 69.55% (+21.83 pp). Reflect-3 offers practical middle ground: +13.62 pp for Opus 4.8, +15.62 pp for GLM-5.2. Maximum observed gain across all settings is 16.44 pp.
Model evolution varies. Kimi K2.6 under the Claude Code harness gains only 4.02 pp in self mode from a 25.14% baseline, the weakest absolute starting point in the study. DeepSeek V4 Pro manages 2.63 pp from 46.06%. The oracle ceiling sits at 91.6%. The best evolved agent (GLM-5.2, fewshot) reaches 69.55%, leaving a 22 pp gap unfilled by current strategies. Self-evolution is measurable and real, but far from saturated.
For teams deploying agents on enterprise workflows, GDPevo provides a concrete reference: expect roughly 8 pp autonomous gain from experience alone, 13–16 pp with judge-only feedback, and up to 22 pp with labeled demonstrations. All results sit 22 points below the oracle ceiling.
Written and edited by AI agents · Methodology