This talk presents methods for uncertainty quantification under partial information, aiming to enable reliable predictions in AI safety and healthcare. At first glance, missing information might seem to make trustworthy predictions impossible. However, I will demonstrate that with appropriate statistical tools, we can achieve reliable predictions even under these challenging conditions.
I will begin by examining safety evaluation for large language models (LLMs). Specifically, we tackle the problem of estimating time-to-unsafe-sampling—how many generations are required before a model produces an unsafe response, such as toxic content.
The core challenge is that unsafe outputs are extremely rare. For many prompts, no unsafe response appears within any feasible number of samples, making direct estimation impractical. This scarcity creates significant obstacles for reliable safety assessment.
Our solution reframes this as a survival analysis problem. We develop a framework that constructs lower bounds on time-to-unsafe-sampling of a given prompt with finite-sample guarantees. By combining conformal prediction with an adaptive, per-prompt sampling strategy, our method delivers formal uncertainty guarantees while efficiently using limited samples. The result is a statistically principled, fine-grained approach to measuring safety risks in generative models.
I will also discuss how conformal prediction can utilize censored data in clinical trials. Real-world medical datasets frequently provide only partial observations—we might know either when a patient experienced an outcome or when they left the study, but rarely both.
I introduce a new conformal prediction framework designed for this general form of censoring. This approach enables valid and informative lower bounds on survival time, providing reliable patient outcome estimates even when datasets have incomplete follow-up information.
In short: partial observations need not prevent trustworthy predictions. Through principled uncertainty quantification, we can transform incomplete data into reliable foresight across critical domains.
