ServiceNow AutoSynthData
ServiceNow AI reported a method for generating verified training tasks from enterprise-agent failures and tested it in two benchmark environments.
What happened lately
ServiceNow AutoSynthData timeline
A failure-guided data pipeline
ServiceNow AI’s AutoSynthData method identifies tasks that a target agent fails but a stronger teacher can solve, then generates new enterprise-workflow tasks around those gaps. Each candidate has a user request, an environment specification and a verifier. The pipeline runs reference solutions and negative checks before accepting a task for training.
The team reports improved mean Pass@1 in the Hybrid and ITSM environments of EnterpriseOps Gym after fine-tuning on generated samples. Those results are limited to the reported benchmark settings. The published report describes the method and links the Gym dataset, but does not announce a generally available AutoSynthData service.