EDGEGEN: Improving Tool-Calling Agents Beyond Happy Paths with Synthetic Edge Case Generation
EdgeGen stress-tests tool-using agents by generating database-grounded tasks meant to break their compliance rules.
The paper says current synthetic task generators often miss an agent’s actual state, database, and edge-case behavior. EdgeGen extracts rules from an agent spec, builds tasks that violate them, then uses the results for finetuning and harness optimization without human annotation. In tau2bench’s airline domain, EdgeGen finetuning improved mean progress by 2% to 42%, while some baselines degraded. For Gemma-4-e4b, its harness optimization beat the human-curated harness by 10% and the base harness by 30%. HF Daily Papers' note
The paper says current synthetic task generators often miss an agent’s actual state, database, and edge-case behavior. EdgeGen extracts rules from an agent spec, builds tasks that violate them, then uses the results for finetuning and harness optimization without human annotation. In tau2bench’s airline domain, EdgeGen finetuning improved mean progress by 2% to 42%, while some baselines degraded. For Gemma-4-e4b, its harness optimization beat the human-curated harness by 10% and the base harness by 30%. HF Daily Papers' note
score 5