Preprint reports gains from training AI agents on generated edge cases

Researchers report that EdgeGen, a system for generating difficult tasks for AI agents, improved performance in an airline-service benchmark. It extracts compliance rules from an agent's specification and creates tasks grounded in a database that test those rules, without requiring people to label the tasks. Fine-tuning with the generated data produced mean progress improvements of 2 percent to 42 percent on the tau2bench airline domain; some baseline methods reduced performance for some models. For one Gemma-4-e4b model, optimizing the agent's test setup improved mean progress by 10 percent over a human-curated setup and 30 percent over the base setup. The reported results are benchmark measurements.