Thanks Devansh for reminding us of the event.
This will be my practice job talk. Feedback on both the research narrative and presentation style is very welcome.
Looking forward to seeing you.
Abstract:
As algorithms increasingly automate critical decisions—from AI-powered text generation to democratic representation—it is more important than ever to design decision-making systems that are both efficient and fair. This talk addresses a fundamental challenge in online decision making: how can we act efficiently and fairly when information arrives sequentially and decisions accumulate over time? I illustrate my research on this question through two recent results. On the efficiency front, I show that the Pandora's Box framework from optimal stopping theory can teach a large language model when to stop generating—reducing inference compute by ~30% with no loss in quality. On the fairness front, I show how principles from computational social choice can strengthen the representativeness of citizen panels selected to discuss policy: while proportionality is achievable in one-shot settings where multiple panels are formed simultaneously, the sequential setting opens the door to fine-grained representation guarantees that grow stronger as decisions accumulate over time. Together, these examples demonstrate how theoretical foundations in online algorithms can address practical challenges in both modern AI systems and democratic institutions.