TRACER: An Innovative Framework for Automatically Exploring and Testing Conversational Agents Using Large Language Models

This article introduces TRACER—an automated testing framework specifically designed for conversational agents. It uses large language models to generate diverse user profiles and test cases, aiming to comprehensively improve the functional coverage and security of chatbots while addressing many challenges faced by traditional testing methods.

In the context of the rapid development of conversational AI, how to efficiently test the functionality and security of chatbots has become a focus of the industry. Traditional testing faces four major challenges:

1. State space explosion: Diverse conversation paths are difficult to cover;
2. Complex intent understanding: User intents are implied in diverse expressions;
3. Hard-to-predict edge cases: Manual enumeration of edge cases and security vulnerabilities is challenging;
4. Personalized interaction needs: Different user profiles require different testing strategies.

TRACER addresses these challenges through three core modules:

• Function Exploration Engine: Uses LLM reasoning capabilities to interact proactively, understand context, and ask exploratory questions to discover hidden functional points;
• User Profile Generator: Automatically generates diverse profiles (different ages/backgrounds, specific goals, edge users, potential malicious users) to ensure testing covers real-world scenarios;
• Test Suite Builder: Generates structured test cases based on exploration results and profiles, covering tests for functionality, process integrity, intent recognition, boundary handling, security, etc.

TRACER's technical highlights include:

1. Adaptive Exploration Strategy: Initial breadth-first discovery of functions, followed by deep digging; LLM adjusts direction based on historical conversations;
2. Multi-dimensional Evaluation System: Covers metrics such as functional coverage, response quality, consistency, and security (e.g., prompt injection, information leakage);
3. Scalable Architecture: Modular design supports integration with different LLM backends and conversational systems; users can customize test parameters (exploration depth, number of profiles, etc.) via configuration.

TRACER has significant value in multiple scenarios:

• Developers: Quickly discover defects and edge cases, evaluate robustness, and perform comprehensive automated testing before release;
• Security Researchers: Systematically find security vulnerabilities, test resistance to adversarial inputs, and evaluate the effectiveness of privacy protection;
• Enterprise Users: Objectively evaluate conversational agent solutions, continuously monitor the performance of deployed systems, and meet compliance testing requirements.

TRACER represents a new paradigm of "AI testing AI". As LLM capabilities improve, using LLMs to test other AI systems will become standard practice, enabling the discovery of issues that traditional testing is hard to capture and adapting to system evolution. In the future, such automated testing frameworks will become a standard part of the conversational agent development process, driving the industry toward higher quality and greater security.