Large Language Models for Program Optimization - Osbert Bastani | Stanford MLSys #91

Episode 91 of the Stanford MLSys Seminar Series!

Large Language Models for Program Optimization
Speaker: Osbert Bastani

Abstract:
Large language models (LLMs) have proven to be surprisingly effective at code generation, making them a promising tool for scaling program synthesis. In this talk, I will describe PIE, a dataset for training LLMs for program optimization based on competitive programming tasks. We benchmark a variety of different techniques using PIE, including in-context learning, chain-of-thought prompting, retrieval-augmented generation, and performance-conditioned finetuning; our strongest models exceed human performance. I will also discuss Eureka, a system that leverages LLMs for synthesizing reward functions in reinforcement learning that outperform human-engineered reward functions. Finally, I will discuss the potential for conformal prediction to improve the trustworthiness of such systems.

Bio:
Osbert Bastani is an assistant professor at the Department of Computer and Information Science at the University of Pennsylvania. He is broadly interested in techniques for designing trustworthy machine learning systems. Previously, he completed his Ph.D. in computer science from Stanford and his A.B. in mathematics from Harvard.

--

Stanford MLSys Seminar hosts: Avanika Narayan, Benjamin Spector, Michael Zhang

Twitter:
  / avanika15​  
  / bfspector  
  / mzhangio  

--

Check out our website for the schedule: http://mlsys.stanford.edu
Join our mailing list to get weekly updates: https://groups.google.com/forum/#!for...

#machinelearning #ai #artificialintelligence #systems #mlsys #computerscience #stanford