Multi-Level Testing of Conversational AI Systems

Technical Summary: Multi-Level Testing of Conversational AI Systems

Problem Statement
Conversational AI systems, increasingly deployed in domains such as healthcare, banking, and education, face significant quality assurance challenges. While users report satisfaction, these systems frequently exhibit critical failures, including the need for users to repeat inputs, the provision of misleading or incorrect information (e.g., hallucinations), and security vulnerabilities like unauthorized access to voice history. Traditional testing strategies are ill-suited for these systems due to the inherent nature of conversational interaction: inputs are natural language sentences with potentially unlimited variations, test oracles must accept semantically equivalent responses, and system behavior involves both responses and external actions. Furthermore, the non-deterministic nature of AI components necessitates probabilistic evaluation rather than binary pass/fail metrics. Existing tools, such as Botium, Charm, and BoTest, offer some capabilities for generating conversational tests or robustness checks but lack the ability to systematically explore the vast conversational space, thoroughly exercise functionalities, or scale effectively to multi-agent architectures.

Methodology
The proposed research adopts a multi-level testing approach, structuring validation across three complementary levels of abstraction to address increasing complexity:

1. Service-Interaction Testing (Level 1): This level focuses on the integration between the language component (e.g., LLMs, NLP pipelines) and the services it invokes. The methodology models test generation as a search problem. It employs feedback-directed gray-box strategies that observe service invocations (e.g., API coverage, parameter coverage) to guide the generation of input sentences. To enhance exploration, the approach integrates LLM-based algorithms to generate diverse, semantically meaningful conversational tests that trigger specific service interactions without requiring full access to internal component details.
2. Agent Testing (Level 2): This level validates whether a single conversational agent fulfills intended goals from the user's perspective. To address the common lack of well-formalized requirements linking conversations to functionalities, the research proposes two strategies:
   • Enriching requirement specifications with conversational data to derive meaningful test cases.
   • Utilizing metamorphic testing, which defines relationships between input transformations and expected output behaviors. This allows for the automatic derivation of test cases even when only sparse knowledge of the expected behavior is available.
3. Multi-Agent System Testing (Level 3): This level assesses the behavior of the system as a whole, where multiple agents coordinate to achieve shared goals. The methodology combines AI planning and orchestration. Planning is used to generate test workflows based on system goals, while orchestration techniques manage the execution of these workflows. Crucially, this level involves injecting special testing and mocking agents into the system to exercise specific behaviors, including erroneous and rare situations, ensuring agents perform the correct actions at the correct times.

Current Progress and Work Plan
The research is structured over three years. Preliminary work has included:

• The creation of a dataset comprising RASA and Dialogflow conversational agents for experimental evaluation.
• A preliminary study on Botium, identifying its limitations.
• The definition of a mutation testing approach for conversational systems to provide fault-based metrics for assessing test suite quality.
• Ongoing development of automatic test case generation for retrieval-based agents using LLMs and the definition of an LLM-driven oracle.

The future work plan involves expanding the dataset to include open-source generative and multi-agent systems. The proposed techniques will be evaluated against baselines (e.g., Botium, Charm) using metrics based on conversational/code coverage, mutation scores, and the detection of real faults.

Key Contributions
The thesis aims to deliver five specific contributions:

1. A curated dataset of conversational AI systems, covering both individual agents and multi-agent systems, to advance quality assurance research.
2. Feedback-directed testing methods to validate language-service interactions and ensure intended functionality.
3. Specification-driven testing methods for individual agents, utilizing requirements and metamorphic relations.
4. Mechanisms for injecting testing and mocking agents into multi-agent systems to validate inter-agent interactions and collaborations.
5. Empirical evidence demonstrating that these methods can reveal faults at different granularity levels, from integration to the system level, thereby enhancing overall system quality.

Significance and Claims
The paper positions this work as a necessary evolution in software engineering for conversational AI. It claims that by moving beyond simple script-based testing to a multi-level, automated approach, the research addresses the "huge, virtually unlimited" conversational space that current tools cannot systematically explore. The significance lies in the ability to detect faults that matter in practice—those developers are willing to fix and practitioners are interested in preventing—across the full spectrum of conversational AI architectures, from single agents to complex collaborative networks. The research does not claim to solve all testing challenges but aims to provide a structured framework that scales with the complexity of modern AI-driven conversational systems.