IAS High Energy Theory Seminar
Solving Quantitative Reasoning Problems with Language Models
The Institute for Advanced Study requires that all adult visitors, collaborators, conference and on-campus seminar attendees and outside vendors coming to the Institute are required to have completed a COVID-19 vaccination and booster in order to enter the IAS campus. Individuals must be prepared to present proof of vaccination if asked and are expected to follow the Institute's Covid-19 Procedures. Masks are optional while indoors.
Additional information can be found at:
https://www.ias.edu/covid-19-updates
This seminar will be presented in-person and on Zoom.
https://theias.zoom.us/j/84116605050?pwd=VHV6VkRUM3hkM2dFSlo2QWJiUWtPdz09
Abstract: Quantitative reasoning tasks which can involve mathematics, science, and programming are often challenging for machine learning models in general and for language models in particular. We show that transformer-based language models obtain significantly better performance on math and science questions when trained in an unsupervised way on a large, math-focused dataset. Performance can be further improved using prompting and sampling techniques including chain-of-thought and majority voting. Minerva, a model that combines these techniques, achieves SOTA on several math and science benchmarks. I will describe the model, its capabilities and limitations.