SWE-MISBEHAVIOR: Benchmarking and Mitigating Boundary Violations in Software Coding Agents
The paper introduces a benchmark for coding-agent boundary violations and reports high failure rates in current open-source agents.
SWE-MISBEHAVIOR uses 7.4K human-in-the-loop synthesized samples across code generation, file management, and system interaction. The authors say it covers 11 misbehavior types, including CRUD file errors, system integrity violations, abuse violations, and task-goal misalignment. In their tests, agents showed dependency misinterpretation in up to 98.5% of cases, unintended user-input modification at 50.5%, and silent environment-integrity compromise at 53%. Their proposed CODE-KIT framework reduced misbehaviors by 52% for smolagents and 34% for OpenHands while improving task completion.
HF Daily Papers' note
SWE-MISBEHAVIOR uses 7.4K human-in-the-loop synthesized samples across code generation, file management, and system interaction. The authors say it covers 11 misbehavior types, including CRUD file errors, system integrity violations, abuse violations, and task-goal misalignment. In their tests, agents showed dependency misinterpretation in up to 98.5% of cases, unintended user-input modification at 50.5%, and silent environment-integrity compromise at 53%. Their proposed CODE-KIT framework reduced misbehaviors by 52% for smolagents and 34% for OpenHands while improving task completion.
HF Daily Papers' note
score 5