Verifiable User Simulation for Search and Recommendation Systems

User simulators lack transparency: Current LLM-based user simulators used to evaluate search engines and recommendation systems are "opaque," making it difficult to understand why simulated users made specific choices or verify those choices match their intended profile—a significant gap when these systems influence real user experiences.

LLMs exhibit demographic bias in simulation: The paper documents that LLMs "can produce biased or discriminatory responses depending on user background characteristics such as language, education level, and cultural context," meaning simulated users from minority or disadvantaged groups may be systematically misrepresented during system evaluation.

A seven-component audit framework enables verification: The proposed framework breaks user simulators into auditable components (Persona, Contract, Execution, Trace, Verification, Feedback, and Refinement), allowing researchers to inspect simulator behavior end-to-end and identify where bias or inconsistencies enter the evaluation process.

Diagnostic analysis differs from statistical validation: The tutorial distinguishes between checking whether a simulator behaves as intended (diagnostic discrepancy analysis) versus whether results are statistically significant, emphasizing that responsible AI evaluation requires both approaches to catch demographic bias that statistics alone might miss.

Responsible AI requires treating simulators as engineered artifacts: By proposing structured design and audit practices for user simulators, the paper argues that evaluation tools themselves must be held to transparency and fairness standards, not just the systems being evaluated.