RESEARCH · RESEARCH · #1423
ServiceNow CoreAI describes AutoSynthData for synthesizing enterprise-agent training tasks
ServiceNow CoreAI presents AutoSynthData, a pipeline that converts a target model's failures into validated training tasks for enterprise agents. It uses evaluation runs in a target environment and a stronger teacher model to generate feasible, realistic, and appropriately difficult tasks (system specification, user prompt, verifier) and iteratively updates the curriculum as the model improves; the paper illustrates the approach with EnterpriseOps Gym and a released dataset (Malay et al., 2026).
KEY POINTS
- ServiceNow CoreAI presents AutoSynthData, a pipeline that converts a target model's failures into validated training tasks for enterprise agents.
- It uses evaluation runs in a target environment and a stronger teacher model to generate feasible, realistic, and appropriately difficult tasks (system specification, user prompt, verifier) and iteratively updates the curriculum as the model improves; the paper illustrates the approach with EnterpriseOps Gym and a released dataset (Malay et al., 2026).
- AutoSynthData provides a practical method to produce grounded, verifiable training examples that target an agent's real-world weaknesses, improving enterprise agent reliability.
WHY IT MATTERS
AutoSynthData provides a practical method to produce grounded, verifiable training examples that target an agent's real-world weaknesses, improving enterprise agent reliability.