Doctoral Speaking Skills Talk - Eungyeup Kim
November 18, 2026 2:00PM—3:00PM
Location:
9115
-
Gates and Hillman Centers
Speaker:
EUNGYEUP KIM,
Ph.D. Student, Computer Science Department, Carnegie Mellon University
https://eungyeupkim.github.io/
While existing benchmarks show near-perfect performance of large language models (LLMs) on many tasks, this apparent saturation obscures the need to rigorously evaluate their reliability. In real-world deployment, each additional "nine": of reliability matters: moving from three nines (99.9\%) to five nines (99.999\%) cuts failures by orders of magnitude, a difference that can be catastrophic in reliability-critical applications. However, estimating such rare failure probabilities with tight confidence bounds requires a prohibitively large number of LLM inferences, making standard Monte Carlo evaluation infeasible under limited compute budgets.
In this project, we observe that across several saturated domains, including easy math reasoning and safety, LLM failures exhibit strong systematic structure: a small subset of inputs accounts for the majority of failures. Leveraging this structure, we use the cross-entropy method (CEM) to learn a sampling distribution concentrated on failure-prone inputs, reducing the number of required inferences by up to two orders of magnitude relative to uniform sampling (e.g., ~156x for GSM8K on Gemini 2.5 Flash Lite and ~10x for sexual-content safety prompts on gpt-oss-20b). This efficiency makes five-nines evaluation computationally feasible, establishing reliability as a distinct and measurable axis of model quality. Our project can be naturally extended to broader domains where failures are rare but costly, such as robotic control and reward hacking in agentic systems.
Presented in Partial Fulfillment of the CSD Speaking Skills Requirement
Contact
Matt Stewart