Georgia Institute of Technology College of Sciences

The success of modern generative models (LLMs) is a technological breakthrough, but foundations on which it is built has remained largely a mystery. A foundational understanding of the distributions that they sample from would vastly improve our ability to both diagnose potential issues (e.g. from a safety perspective) and also ability to push the applications of these models (e.g. in challenging settings such as scientific discovery). In this talk, I will broadly survey some aspects of work in this line while zooming in on a particular aspect: understanding sampling at inference time. In particular, we ask the question of how we can elicit, understand and control new behavior from language models particularly aimed towards long horizon tasks such as reasoning. Typical algorithmic interventions here suffer the issue of having errors compound with the time horizon. In this talk, we will see how these questions here can be connected to classical algorithmic questions from the theory of sampling and Markov chain Monte Carlo.  Using this new perspective, we will present an algorithm VGB, that circumvents compounding of error with the horizon. Time permitting, we will survey the broader algorithmic landscape, with further connections to particle filtering and stochastic control. 

This talk is based on joint work with Dhruv Rohatgi, Yuchen Li, Donya Salless, Andrej Risteski, Ankur Moitra and Dylan Foster.