Leonard Tang is co-founder and chief executive of Haize Labs, which develops tools for testing, evaluating, and improving generative AI systems before they enter production. His central concern is generative-AI brittleness: small changes in wording, context, or audio can produce dramatically different behavior, even when an application passes its existing tests.
Tang studied mathematics and computer science at Harvard, researching adversarial robustness, mathematical reasoning, computational neuroscience, interpretability, and language models. After admission to Stanford’s computer-science doctoral program, he chose entrepreneurship instead. He founded Haize Labs in 2023 with Richard Liu and Steve Li, initially focusing on jailbreaks, hallucinations, and other vulnerabilities.
Haizing: Tang applies property-based testing and fuzzing to generative systems by simulating varied inputs, scoring responses against application-specific requirements, and directing subsequent searches toward likely failures. His AI Engineer World’s Fair presentation describes tree search, embedding-space exploration, and gradient-based optimization for testing adversarial prompts, multistep conversations, and noisy voice interactions.
Verdict: With Nimit Kalra, Tang created Verdict, an open-source framework for composing more dependable AI judges from debate, self-verification, ensembles, and detailed rubrics. Their research on scaling judge-time compute investigates whether structured evaluation workflows can improve accuracy and interpretability without relying solely on larger models.
Reward models for subjective work: Tang explores reinforcement-learning-trained evaluators that generate task-specific criteria and explain their assessments. The objective is to convert expert judgment into feedback that models can use when correctness cannot be checked with a simple test.
His work extends to financial services and customer-facing voice agents, where realistic evaluation must account for prompt injection, conversational history, background noise, and the operational consequences of unreliable responses.
