EdgeGen Generates Rule-Violating Edge-Case Tasks From an Agent's Spec and Improves τ²-bench Airline Progress by 2-42% With No Human Labels

arXiv·medium signalXBlueskyLinkedInCopy link

EdgeGen extracts compliance rules from an agent's specification and generates database-grounded tasks designed to violate them, targeting the unhappy paths that generic synthetic data misses. Fine-tuning on this data improved mean progress on τ²-bench airline by 2 to 42 percent, while baseline synthetic methods made some models worse. Used for harness optimization on Gemma-4-e4b, it beat the human-curated harness by 10 percent and the base harness by 30 percent. The same method could drive regression suites for any tool-calling agent that has a written policy.