DiCriTest: Testing Scenario Generation for Decision-Making Agents Considering Diversity and Criticality

The growing deployment of decision-making agents in dynamic environments increases the demand for safety verification. While critical testing scenario generation has emerged as an appealing verification methodology, effectively balancing diversity and criticality remains a key challenge for existing methods, particularly due to local optima in high-dimensional scenario spaces. To address this limitation, we propose a dual-space guided testing framework that coordinates the scenario parameter space and the agent behavior space, aiming to generate testing scenarios considering diversity and criticality. Specifically, in the scenario parameter space, a hierarchical representation framework combines dimensionality reduction and multi-dimensional cube evaluation to efficiently localize diverse and critical cubes. This guides dynamic coordination between two generation modes: local perturbation and global exploration, optimizing critical scenario quantity and diversity. Complementarily, in the agent behavior space, agent-environment interaction data are leveraged to quantify behavioral criticality/diversity and adaptively support generation mode switching, forming a closed feedback loop that continuously enhances scenario characterization and exploration within the parameter space. Experiments show our framework improves critical scenario generation by an average of 71.37% and demonstrates greater diversity under novel parameter-behavior co-driven metrics when tested on five decision-making agents, outperforming state-of-the-art baselines. Note to Practitioners - Ensuring the safety of decision-making agents in complex real-world environments demands rigorous testing under diverse, realistic conditions. A major practical challenge in automated testing is efficiently uncovering both severe failure scenarios and a broad variety of failure types. Existing methods often get trapped in repetitive cycles, generating highly similar test scenarios that fail to cover broader potential risks. DiCriTest addresses this by intelligently guiding testing scenario generation through two complementary perspectives: proactively expanding the scope of physical scenario parameters while simultaneously capturing varied decision behaviors during agent-environment interactions. This dual approach allows the framework to dynamically adjust between intensively probing high-risk areas and exploring new scenario types, significantly expanding test coverage. Compared to existing methods, DiCriTest detects over 70% more critical scenarios on average while better capturing real-world risk diversity. Designed as a general-purpose testing method for intelligent agents, it requires no domain-specific expertise and can be readily applied to test diverse decision-making agents.